prithivMLmods

The Deep-Fake-Detector-v2-Model is a state-of-the-art deep learning model designed to detect deepfake images. It leverages the Vision Transformer (ViT) architecture, specifically the `google/vit-base-patch16-224-in21k` model, fine-tuned on a dataset of real and deepfake images. The model is trained to classify images as either "Realism" or "Deepfake" with high accuracy, making it a powerful tool for detecting manipulated media. Update : The previous model checkpoint was obtained using a smaller classification dataset. Although it performed well in evaluation scores, its real-time performance was average due to limited variations in the training set. The new update includes a larger dataset to improve the detection of fake images. | Repository | Link | |------------|------| | Deep Fake Detector v2 Model | GitHub Repository | Key Features - Architecture: Vision Transformer (ViT) - `google/vit-base-patch16-224-in21k`. - Input: RGB images resized to 224x224 pixels. - Output: Binary classification ("Realism" or "Deepfake"). - Training Dataset: A curated dataset of real and deepfake images. - Fine-Tuning: The model is fine-tuned using Hugging Face's `Trainer` API with advanced data augmentation techniques. - Performance: Achieves high accuracy and F1 score on validation and test datasets. Model Architecture The model is based on the Vision Transformer (ViT), which treats images as sequences of patches and applies a transformer encoder to learn spatial relationships. Key components include: - Patch Embedding: Divides the input image into fixed-size patches (16x16 pixels). - Transformer Encoder: Processes patch embeddings using multi-head self-attention mechanisms. - Classification Head: A fully connected layer for binary classification. Training Details - Optimizer: AdamW with a learning rate of `1e-6`. - Batch Size: 32 for training, 8 for evaluation. - Epochs: 2. - Data Augmentation: - Random rotation (±90 degrees). - Random sharpness adjustment. - Random resizing and cropping. - Loss Function: Cross-Entropy Loss. - Evaluation Metrics: Accuracy, F1 Score, and Confusion Matrix. Dataset The model is fine-tuned on the dataset, which contains: - Real Images: Authentic images of human faces. - Fake Images: Deepfake images generated using advanced AI techniques. Limitations The model is trained on a specific dataset and may not generalize well to other deepfake datasets or domains. - Performance may degrade on low-resolution or heavily compressed images. - The model is designed for image classification and does not detect deepfake videos directly. Misuse: This model should not be used for malicious purposes, such as creating or spreading deepfakes. Bias: The model may inherit biases from the training dataset. Care should be taken to ensure fairness and inclusivity. Transparency: Users should be informed when deepfake detection tools are used to analyze their content. Future Work - Extend the model to detect deepfake videos. - Improve generalization by training on larger and more diverse datasets. - Incorporate explainability techniques to provide insights into model predictions. ```bibtex @misc{Deep-Fake-Detector-v2-Model, author = {prithivMLmods}, title = {Deep-Fake-Detector-v2-Model}, initial = {21 Mar 2024}, secondupdated = {31 Jan 2025}, latestupdated = {02 Feb 2025} }

> Watermark-Detection-SigLIP2 is a vision-language encoder model fine-tuned from google/siglip2-base-patch16-224 for binary image classification. It is trained to detect whether an image contains a watermark or not, using the SiglipForImageClassification architecture. > [!note] > Watermark detection works best with crisp and high-quality images. Noisy images are not recommended for validation. > [!note] SigLIP 2: Multilingual Vision-Language Encoders with Improved Semantic Understanding, Localization, and Dense Features https://arxiv.org/pdf/2502.14786 Watermark-Detection-SigLIP2 is useful in scenarios such as: - Content Moderation – Automatically detect watermarked content on image sharing platforms. - Dataset Cleaning – Filter out watermarked images from training datasets. - Copyright Enforcement – Monitor and flag usage of watermarked media. - Digital Forensics – Support analysis of tampered or protected media assets.

> The Qwen2-VL-OCR-2B-Instruct model is a fine-tuned version of Qwen/Qwen2-VL-2B-Instruct, tailored for tasks that involve Optical Character Recognition (OCR), image-to-text conversion, and math problem solving with LaTeX formatting. This model integrates a conversational approach with visual and textual understanding to handle multi-modal tasks effectively. [](https://huggingface.co/prithivMLmods/Qwen2-VL-OCR-2B-Instruct/blob/main/Demo/ocrtestqwen.ipynb) SoTA understanding of images of various resolution & ratio: Qwen2-VL achieves state-of-the-art performance on visual understanding benchmarks, including MathVista, DocVQA, RealWorldQA, MTVQA, etc. Understanding videos of 20min+: Qwen2-VL can understand videos over 20 minutes for high-quality video-based question answering, dialog, content creation, etc. Agent that can operate your mobiles, robots, etc.: with the abilities of complex reasoning and decision making, Qwen2-VL can be integrated with devices like mobile phones, robots, etc., for automatic operation based on visual environment and text instructions. Multilingual Support: to serve global users, besides English and Chinese, Qwen2-VL now supports the understanding of texts in different languages inside images, including most European languages, Japanese, Korean, Arabic, Vietnamese, etc. | File Name | Size | Description | Upload Status | |---------------------------|------------|------------------------------------------------|-------------------| | `.gitattributes` | 1.52 kB | Configures LFS tracking for specific model files. | Initial commit | | `README.md` | 203 Bytes | Minimal details about the uploaded model. | Updated | | `addedtokens.json` | 408 Bytes | Additional tokens used by the model tokenizer. | Uploaded | | `chattemplate.json` | 1.05 kB | Template for chat-based model input/output. | Uploaded | | `config.json` | 1.24 kB | Model configuration metadata. | Uploaded | | `generationconfig.json` | 252 Bytes | Configuration for text generation settings. | Uploaded | | `merges.txt` | 1.82 MB | BPE merge rules for tokenization. | Uploaded | | `model.safetensors` | 4.42 GB | Serialized model weights in a secure format. | Uploaded (LFS) | | `preprocessorconfig.json`| 596 Bytes | Preprocessing configuration for input data. | Uploaded | | `vocab.json` | 2.78 MB | Vocabulary file for tokenization. | Uploaded | 1. Vision-Language Integration: - Combines image understanding with natural language processing to convert images into text. 2. Optical Character Recognition (OCR): - Extracts and processes textual information from images with high accuracy. 3. Math and LaTeX Support: - Solves math problems and outputs equations in LaTeX format. 4. Conversational Capabilities: - Designed to handle multi-turn interactions, providing context-aware responses. 5. Image-Text-to-Text Generation: - Inputs can include images, text, or a combination, and the model generates descriptive or problem-solving text. 6. Secure Weight Format: - Uses Safetensors for faster and more secure model weight loading. - Base Model: Qwen/Qwen2-VL-2B-Instruct - Model Size: - 2.21 Billion parameters - Optimized for BF16 tensor type, enabling efficient inference. - Specializations: - OCR tasks in images containing text. - Mathematical reasoning and LaTeX output for equations.

> open-scene-detection is a vision-language encoder model fine-tuned from `siglip2-base-patch16-512` for multi-class scene classification. It is trained to recognize and categorize natural and urban scenes using a curated visual dataset. The model uses the `SiglipForImageClassification` architecture. The model classifies an image into one of the following scenes: Scene Recognition – Automatically classify natural and urban scenes. Environmental Mapping – Support geographic and ecological analysis from visual data. Dataset Annotation – Efficiently label large-scale image datasets by scene. Visual Search and Organization – Enable smart scene-based filtering or retrieval. Autonomous Systems – Assist navigation and perception modules with scene understanding.

> Qwen3-VL-4B-Thinking-abliterated is an abliterated (v1.0) variant of Qwen3-VL-4B-Thinking, designed for Abliterated Reasoning and Captioning. This model generates detailed captions and reasoning outputs across a wide range of visual and multimodal contexts, including complex, sensitive, or nuanced content, and supports diverse aspect ratios and resolutions. Abliterated / Uncensored Captioning: Fine-tuned to bypass standard content filters while preserving factual, descriptive, and reasoning-rich outputs. High-Fidelity Descriptions: Produces comprehensive captions and reasoning for general, artistic, technical, abstract, or low-context images. Robust Across Aspect Ratios: Supports wide, tall, square, and irregular image dimensions with consistent accuracy. Variational Detail Control: Generates outputs ranging from high-level summaries to fine-grained, intricate descriptions and reasoning. Foundation on Qwen3-VL-4B-Thinking Architecture: Leverages Qwen3-VL-4B-Thinking’s multimodal reasoning and instruction-following capabilities. Multilingual Output Capability: Primarily English, with adaptability for multilingual prompts via prompt engineering. Generating detailed, uncensored captions and reasoning for general-purpose or artistic datasets. Research in content moderation, red-teaming, and generative safety evaluation. Enabling descriptive captioning and reasoning for visual datasets typically excluded from mainstream models. Creative applications such as storytelling, art generation, or multimodal reasoning tasks. Captioning and reasoning for non-standard aspect ratios and stylized visual content. May produce explicit, sensitive, or offensive descriptions depending on image content and prompts. Not recommended for production systems requiring strict content moderation. Output style, tone, and reasoning can vary depending on input phrasing. Accuracy may vary for unfamiliar, synthetic, or highly abstract visual content.

> Qwen3-VL-4B-Instruct-abliterated is an abliterated (v1.0) variant of Qwen3-VL-4B-Instruct, tailored for Abliterated Reasoning and Captioning. This model is designed to generate detailed and descriptive captions, as well as reasoning outputs, across a wide range of visual and multimodal contexts—including complex, sensitive, or nuanced content—while supporting diverse aspect ratios and resolutions. Abliterated / Uncensored Captioning: Fine-tuned to bypass conventional content filters while preserving factual, descriptive, and reasoning-rich outputs. High-Fidelity Descriptions: Generates comprehensive captions and reasoning for general, artistic, technical, abstract, or low-context images. Robust Across Aspect Ratios: Supports wide, tall, square, and irregular image dimensions with consistent accuracy. Variational Detail Control: Produces outputs ranging from high-level summaries to fine-grained, intricate descriptions and reasoning. Foundation on Qwen3-VL-4B Architecture: Leverages Qwen3-VL-4B’s multimodal reasoning and instruction-following capabilities. Multilingual Output Capability: Primarily English, with adaptability for multilingual prompts via prompt engineering. Generating detailed, uncensored captions and reasoning for general-purpose or artistic datasets. Research in content moderation, red-teaming, and generative safety evaluation. Enabling descriptive captioning and reasoning for visual datasets typically excluded from mainstream models. Creative applications such as storytelling, art generation, or multimodal reasoning tasks. Captioning and reasoning for non-standard aspect ratios and stylized visual content. May produce explicit, sensitive, or offensive descriptions depending on image content and prompts. Not recommended for production systems requiring strict content moderation. Output style, tone, and reasoning can vary depending on input phrasing. Accuracy may vary for unfamiliar, synthetic, or highly abstract visual content.

> The Huihui-Qwen3-4B-Instruct-2507-abliterated model is an uncensored, proof-of-concept version of the Qwen3-4B-Instruct-2507 large language model, created using a novel abliteration method designed to remove refusal responses without using TransformerLens. This approach offers a faster and more effective way to bypass the model's standard refusal behaviors, resulting in a less filtered and more raw output experience, though it lacks rigorous safety filtering and may generate sensitive or controversial content. The model is quantized (4-bit) for efficient use, can be used directly in Hugging Face’s transformers library, and is intended primarily for research or experimental use rather than production due to the reduced content restrictions and associated risks. Users are advised to carefully monitor outputs and ensure ethical and legal compliance when deploying this model. | Model Variant | Link | |--------------|------| | Qwen3-4B-Thinking-2507-abliterated-GGUF | Hugging Face | | Qwen3-4B-Instruct-2507-abliterated-GGUF | Hugging Face | | File Name | Size | Quant Type | |-----------|------|------------| | Qwen3-4B-Thinking-2507-abliterated.BF16.gguf | 8.05 GB | BF16 | | Qwen3-4B-Thinking-2507-abliterated.F16.gguf | 8.05 GB | F16 | | Qwen3-4B-Thinking-2507-abliterated.F32.gguf | 16.1 GB | F32 | | Qwen3-4B-Thinking-2507-abliterated.Q2K.gguf | 1.67 GB | Q2K | | Qwen3-4B-Thinking-2507-abliterated.Q3KL.gguf | 2.24 GB | Q3KL | | Qwen3-4B-Thinking-2507-abliterated.Q3KM.gguf | 2.08 GB | Q3KM | | Qwen3-4B-Thinking-2507-abliterated.Q3KS.gguf | 1.89 GB | Q3KS | | Qwen3-4B-Thinking-2507-abliterated.Q4KM.gguf | 2.5 GB | Q4KM | | Qwen3-4B-Thinking-2507-abliterated.Q4KS.gguf | 2.38 GB | Q4KS | | Qwen3-4B-Thinking-2507-abliterated.Q5KM.gguf | 2.89 GB | Q5KM | | Qwen3-4B-Thinking-2507-abliterated.Q5KS.gguf | 2.82 GB | Q5KS | | Qwen3-4B-Thinking-2507-abliterated.Q6K.gguf | 3.31 GB | Q6K | | Qwen3-4B-Thinking-2507-abliterated.Q80.gguf | 4.28 GB | Q80 | | File Name | Size | Quant Type | |-----------|------|------------| | Qwen3-4B-Instruct-2507-abliterated.BF16.gguf | 8.05 GB | BF16 | | Qwen3-4B-Instruct-2507-abliterated.F16.gguf | 8.05 GB | F16 | | Qwen3-4B-Instruct-2507-abliterated.F32.gguf | 16.1 GB | F32 | | Qwen3-4B-Instruct-2507-abliterated.Q2K.gguf | 1.67 GB | Q2K | | Qwen3-4B-Instruct-2507-abliterated.Q3KL.gguf | 2.24 GB | Q3KL | | Qwen3-4B-Instruct-2507-abliterated.Q3KM.gguf | 2.08 GB | Q3KM | | Qwen3-4B-Instruct-2507-abliterated.Q3KS.gguf | 1.89 GB | Q3KS | | Qwen3-4B-Instruct-2507-abliterated.Q4KM.gguf | 2.5 GB | Q4KM | | Qwen3-4B-Instruct-2507-abliterated.Q4KS.gguf | 2.38 GB | Q4KS | | Qwen3-4B-Instruct-2507-abliterated.Q5KM.gguf | 2.89 GB | Q5KM | | Qwen3-4B-Instruct-2507-abliterated.Q5KS.gguf | 2.82 GB | Q5KS | | Qwen3-4B-Instruct-2507-abliterated.Q6K.gguf | 3.31 GB | Q6K | | Qwen3-4B-Instruct-2507-abliterated.Q80.gguf | 4.28 GB | Q80 | (sorted by size, not necessarily quality. IQ-quants are often preferable over similar sized non-IQ quants) Here is a handy graph by ikawrakow comparing some lower-quality quant types (lower is better):

> Age-Classification-SigLIP2 is an image classification vision-language encoder model fine-tuned from google/siglip2-base-patch16-224 for a single-label classification task. It is designed to predict the age group of a person from an image using the SiglipForImageClassification architecture. The model categorizes images into five age groups: - Class 0: "Child 0-12" - Class 1: "Teenager 13-20" - Class 2: "Adult 21-44" - Class 3: "Middle Age 45-64" - Class 4: "Aged 65+" The Age-Classification-SigLIP2 model is designed to classify images into five age categories. Potential use cases include: - Demographic Analysis: Helping businesses and researchers analyze age distribution. - Health & Fitness Applications: Assisting in age-based health recommendations. - Security & Access Control: Implementing age verification in digital systems. - Retail & Marketing: Enhancing personalized customer experiences. - Forensics & Surveillance: Aiding in age estimation for security purposes.

> DeepSeek-OCR-Latest-BF16.I64 is an optimized and updated version of the original DeepSeek-OCR. It is an open-source vision-language OCR model designed to extract text from images and scanned documents—including both digital and handwritten content—and can output results as plain text or Markdown. This model leverages a powerful multimodal backbone (3B VLM) to improve reading comprehension and layout understanding for both typed and cursive handwriting. It also excels at preserving document structures such as headings, tables, and lists in its outputs. The BF16 variant has been updated and tested with the following environment: This version allows flexible configuration of attention implementations—such as `flashattention` or `sdpa`—for performance optimization or standardization. Users can also opt out of specific attention implementations if desired. | Resource Type | Description | Link | |----------------|--------------|------| | Original Model Card | Official DeepSeek-OCR release by deepseek-ai | deepseek-ai/DeepSeek-OCR | | Test Model (StrangerZone HF) | Community test deployment (experimental) | strangervisionhf/deepseek-ocr-latest-transformers | | Standard Model Card | Optimized version supporting Transformers v4.57.1 (BF16 precision) | DeepSeek-OCR-Latest-BF16.I64 | | Research Paper | DeepSeek-OCR: Contexts Optical Compression | arXiv:2510.18234 | | Demo Space | Interactive demo hosted on Hugging Face Spaces | DeepSeek-OCR Experimental Demo |

- Hosted Here🧨: https://huggingface.co/spaces/prithivMLmods/FLUX-LoRA-DLC The model is still in the training phase. This is not the final version and may contain artifacts and perform poorly in some cases. | Parameter | Value | Parameter | Value | |---------------------------|--------|---------------------------|--------| | LR Scheduler | constant | Noise Offset | 0.03 | | Optimizer | AdamW | Multires Noise Discount | 0.1 | | Network Dim | 64 | Multires Noise Iterations | 10 | | Network Alpha | 32 | Repeat & Steps | 24 & 2340| | Epoch | 15 | Save Every N Epochs | 1 | You should use `Retro Pixel` to trigger the image generation. Weights for this model are available in Safetensors format.

> security-llama3.2-3b is a dense, decoder-only Transformer model with approximately 3 billion parameters. It is optimized for generating text, particularly in response to prompts in a chat format, with a context length of up to 4,000 tokens. The model is specialized toward cybersecurity content, drawing on a mixture of publicly available blogs, papers, reference datasets (e.g. from the PEASEC cybersecurity repository), synthetic “textbook-style” data, and academic Q&A sources to enhance performance in security-themed tasks. For usage, the model accepts chat-style inputs (e.g. alternating “user” / “assistant” messages) and can be deployed via the Hugging Face transformers library (e.g. via pipeline("text-generation", model="viettelsecurity-ai/security-llama3.2-3b")). The model weights are stored in safetensors format, configured with fp16 (half precision), and no inference provider currently hosts it by default. | File Name | Quant Type | File Size | | - | - | - | | security-llama3.2-3b.BF16.gguf | BF16 | 6.43 GB | | security-llama3.2-3b.F16.gguf | F16 | 6.43 GB | | security-llama3.2-3b.F32.gguf | F32 | 12.9 GB | | security-llama3.2-3b.Q2K.gguf | Q2K | 1.36 GB | | security-llama3.2-3b.Q3KL.gguf | Q3KL | 1.82 GB | | security-llama3.2-3b.Q3KM.gguf | Q3KM | 1.69 GB | | security-llama3.2-3b.Q3KS.gguf | Q3KS | 1.54 GB | | security-llama3.2-3b.Q40.gguf | Q40 | 1.92 GB | | security-llama3.2-3b.Q41.gguf | Q41 | 2.09 GB | | security-llama3.2-3b.Q4K.gguf | Q4K | 2.02 GB | | security-llama3.2-3b.Q4KM.gguf | Q4KM | 2.02 GB | | security-llama3.2-3b.Q4KS.gguf | Q4KS | 1.93 GB | | security-llama3.2-3b.Q50.gguf | Q50 | 2.27 GB | | security-llama3.2-3b.Q51.gguf | Q51 | 2.45 GB | | security-llama3.2-3b.Q5K.gguf | Q5K | 2.32 GB | | security-llama3.2-3b.Q5KM.gguf | Q5KM | 2.32 GB | | security-llama3.2-3b.Q5KS.gguf | Q5KS | 2.27 GB | | security-llama3.2-3b.Q6K.gguf | Q6K | 2.64 GB | | security-llama3.2-3b.Q80.gguf | Q80 | 3.42 GB | (sorted by size, not necessarily quality. IQ-quants are often preferable over similar sized non-IQ quants) Here is a handy graph by ikawrakow comparing some lower-quality quant types (lower is better):

> The Gliese-OCR-7B-Post2.0-final model is a refined and optimized version of Gliese-OCR-7B-Post1.0, built upon the Qwen2.5-VL architecture. It represents the final iteration in the Gliese-OCR series, offering enhanced efficiency, precision, and visualization capabilities for document OCR, visual analysis, and information extraction. > > Fine-tuned with extended document visualization data and OCR-focused objectives, this model delivers superior accuracy across a wide range of document types, including scanned PDFs, handwritten pages, structured forms, and analytical reports. Optimized Document Visualization and OCR Pipeline: Significantly improved recognition of text, layout, and embedded visuals for structured document understanding. Context-Aware Multimodal Linking: Enhanced understanding of document context with stronger alignment between text, images, and layout components. Refined Document Retrieval: Improved retrieval accuracy from complex layouts and multi-page documents. High-Fidelity Content Extraction: Precise extraction of structured, semi-structured, and unstructured information with advanced text normalization. Analytical Recognition: Superior reasoning over charts, graphs, tables, and mathematical equations. Improved Visual Reasoning and Layout Awareness: Trained on document visualization datasets for advanced spatial and semantic comprehension. State-of-the-Art Performance Across Resolutions: Achieves top results on benchmarks such as DocVQA, InfographicVQA, MathVista, and RealWorldQA. Extended Multimodal Duration Support: Handles long document sequences and extended videos (20+ minutes). Final Release Stability: Consolidates all prior improvements for stable and reliable performance. Document visualization and OCR extraction tasks. Context-aware document retrieval and multimodal linking. Extraction and LaTeX formatting of equations and structured content. Analytical document interpretation (charts, tables, graphs, and figures). Multilingual OCR for enterprise, academic, and research use cases. Summarization, question answering, and cross-modal reasoning over long documents. Intelligent robotic or mobile automation guided by visual document input. Reduced accuracy on heavily degraded or occluded documents. High computational requirements for large-scale or real-time applications. Limited optimization for low-resource or edge devices. Occasional misalignment in text layout or minor hallucinations in outputs. Performance may vary depending on visual token configuration and context length settings.

> The Qwen2.5-VL-7B-Abliterated-Caption-it model is a fine-tuned version of Qwen2.5-VL-7B-Instruct, tailored for Abliterated Captioning / Uncensored Image Captioning. This variant is designed to generate highly detailed and descriptive captions across a broad range of visual categories, including images with complex, sensitive, or nuanced content—across varying aspect ratios and resolutions. Abliterated / Uncensored Captioning: Fine-tuned to bypass common content filters while preserving factual and descriptive richness across diverse visual categories. High-Fidelity Descriptions: Generates comprehensive captions for general, artistic, technical, abstract, and low-context images. Robust Across Aspect Ratios: Capable of accurately captioning images with wide, tall, square, and irregular dimensions. Variational Detail Control: Produces outputs with both high-level summaries and fine-grained descriptions as needed. Foundation on Qwen2.5-VL Architecture: Leverages the strengths of the Qwen2.5-VL-7B multimodal model for visual reasoning, comprehension, and instruction-following. Multilingual Output Capability: Can support multilingual descriptions (English as default), adaptable via prompt engineering. This model was fine-tuned using the following datasets: prithivMLmods/blip3o-caption-mini-arrow prithivMLmods/Caption3o-Opt-v2 Private/unlisted datasets curated for uncensored and domain-specific image captioning tasks. The training objective focused on enhancing performance in unconstrained, descriptive image captioning—especially for edge cases commonly filtered out in standard captioning benchmarks. > [!note] Instruction Query: Provide a detailed caption for the image Generating detailed and unfiltered image captions for general-purpose or artistic datasets. Content moderation research, red-teaming, and generative safety evaluations. Enabling descriptive captioning for visual datasets typically excluded from mainstream models. Use in creative applications (e.g., storytelling, art generation) that benefit from rich descriptive captions. Captioning for non-standard aspect ratios and stylized visual content. May produce explicit, sensitive, or offensive descriptions depending on image content and prompts. Not suitable for deployment in production systems requiring content filtering or moderation. Can exhibit variability in caption tone or style depending on input prompt phrasing. Accuracy for unfamiliar or synthetic visual styles may vary.

> Food-101-93M is a fine-tuned image classification model built on top of google/siglip2-base-patch16-224 using the SiglipForImageClassification architecture. It is trained to classify food images into one of 101 popular dishes, derived from the Food-101 dataset. The model categorizes images into 101 food classes such as `sushi`, `hamburger`, `waffles`, `padthai`, and more. - Recipe Recommendation Engines: Automatically tagging food images to suggest recipes. - Food Logging & Calorie Tracking Apps: Categorizing meals based on photos. - Smart Kitchens: Assisting food recognition in smart appliances. - Restaurant Menu Digitization: Auto-classifying dishes for visual menus or ordering systems. - Dataset Labeling: Enabling automatic annotation of food datasets for training other ML models.

> The Gliese-OCR-7B-Post1.0 model is a fine-tuned version of Camel-Doc-OCR-062825, optimized for Document Retrieval, Content Extraction, and Analysis Recognition. Built on top of the Qwen2.5-VL architecture, this model enhances document comprehension capabilities with focused training on the Opendoc2-Analysis-Recognition dataset for superior document analysis and information extraction tasks. > [!note] This model shows significant improvements in LaTeX rendering and Markdown rendering for OCR tasks-reportlab/GlieseOCR7BPost10(4bit)reportlab.ipynb). Context-Aware Multimodal Extraction and Linking for Documents: Advanced capability for understanding document context and establishing connections between multimodal elements within documents. Enhanced Document Retrieval: Designed to efficiently locate and extract relevant information from complex document structures and layouts. Superior Content Extraction: Optimized for precise extraction of structured and unstructured content from diverse document formats. Analysis Recognition: Specialized in recognizing and interpreting analytical content, charts, tables, and visual data representations. State-of-the-Art Performance Across Resolutions: Achieves competitive results on OCR and visual QA benchmarks such as DocVQA, MathVista, RealWorldQA, and MTVQA. Video Understanding up to 20+ minutes: Supports detailed comprehension of long-duration videos for content summarization, Q\&A, and multi-modal reasoning. Visually-Grounded Device Interaction: Enables mobile/robotic device operation via visual inputs and text-based instructions using contextual understanding and decision-making logic. Context-aware multimodal extraction and linking for complex document structures. High-fidelity document retrieval and content extraction from various document formats. Analysis recognition of charts, graphs, tables, and visual data representations. Document-based question answering for educational and enterprise applications. Extraction and LaTeX formatting of mathematical expressions from printed or handwritten content. Retrieval and summarization from long documents, slides, and multi-modal inputs. Multilingual document analysis and structured content extraction for global use cases. Robotic or mobile automation with vision-guided contextual interaction. May show degraded performance on extremely low-quality or occluded images. Not optimized for real-time applications on low-resource or edge devices due to computational demands. Variable accuracy on uncommon or low-resource languages/scripts. Long video processing may require substantial memory and is not optimized for streaming applications. Visual token settings affect performance; suboptimal configurations can impact results. In rare cases, outputs may contain hallucinated or contextually misaligned information.

| Parameter | Value | Parameter | Value | |---------------------------|--------|---------------------------|--------| | LR Scheduler | constant | Noise Offset | 0.03 | | Optimizer | AdamW | Multires Noise Discount | 0.1 | | Network Dim | 64 | Multires Noise Iterations | 10 | | Network Alpha | 32 | Repeat & Steps | 20 & 2790 | | Epoch | 20 | Save Every N Epochs | 1 | | Source | Link | |--------------|-------------------------------------| | Playground | playground.com | | ArtStation | artstation.com | | 4K Wallpapers| 4kwallpapers.com | | Dimensions | Aspect Ratio | Recommendation | |-----------------|------------------|---------------------------| | 1472 x 1140 | 4:3 (approx.) | Best | | 1024 x 1024 | 1:1 | Default | You should use `Studio Realism` to trigger the image generation.

> II-Search-4B is a 4-billion-parameter language model fine-tuned from Qwen3-4B specifically for advanced information seeking and web-integrated reasoning tasks, demonstrating strong capabilities in multi-hop information retrieval, fact verification, and comprehensive report generation; it excels on factual QA benchmarks compared to peers, features sophisticated tool-use for search and web visits, supports distributed inference with vLLM or SGLang (including a 131,072-token context window with custom RoPE scaling), and is suitable for factual question answering, research assistance, and educational applications, with Apple Silicon support via MLX, open integration examples, and full resources available on its Hugging Face repository. | File Name | Size | Quant Type | |-----------|------|------------| | II-Search-4B-GGUF.BF16.gguf | 8.05 GB | BF16 | | II-Search-4B-GGUF.F16.gguf | 8.05 GB | F16 | | II-Search-4B-GGUF.F32.gguf | 16.1 GB | F32 | | II-Search-4B-GGUF.Q2K.gguf | 1.67 GB | Q2K | | II-Search-4B-GGUF.Q3KL.gguf | 2.24 GB | Q3KL | | II-Search-4B-GGUF.Q3KM.gguf | 2.08 GB | Q3KM | | II-Search-4B-GGUF.Q3KS.gguf | 1.89 GB | Q3KS | | II-Search-4B-GGUF.Q4KM.gguf | 2.5 GB | Q4KM | | II-Search-4B-GGUF.Q4KS.gguf | 2.38 GB | Q4KS | | II-Search-4B-GGUF.Q5KM.gguf | 2.89 GB | Q5KM | | II-Search-4B-GGUF.Q5KS.gguf | 2.82 GB | Q5KS | | II-Search-4B-GGUF.Q6K.gguf | 3.31 GB | Q6K | | II-Search-4B-GGUF.Q80.gguf | 4.28 GB | Q80 | (sorted by size, not necessarily quality. IQ-quants are often preferable over similar sized non-IQ quants) Here is a handy graph by ikawrakow comparing some lower-quality quant types (lower is better):

Photo-Restore-i2i is an adapter for black-forest-lab's FLUX.1-Kontext-dev, designed to restore old photos into mid-colorized, detailed images. The model was trained on 50 image pairs (25 start images, 25 end images). Synthetic result nodes were generated using NanoBanana from Google and SeedDream 4 (dataset for result sets), and labeled with DeepCaption-VLA-7B. The adapter is triggered with the following prompt: > [!note] [photo content], restore and enhance the image by repairing any damage, scratches, or fading. Colorize the photo naturally while preserving authentic textures and details, maintaining a realistic and historically accurate look. | Setting | Value | | ------------------------ | ------------------------ | | Module Type | Adapter | | Base Model | FLUX.1 Kontext Dev - fp8 | | Trigger Words | [photo content], restore and enhance the image by repairing any damage, scratches, or fading. Colorize the photo naturally while preserving authentic textures and details, maintaining a realistic and historically accurate look. | | Image Processing Repeats | 50 | | Epochs | 28 | | Save Every N Epochs | 1 | Labeling: DeepCaption-VLA-7B(natural language & English) Total Images Used for Training : 50 Image Pairs (25 Start, 25 End) Synthetic Result Node generated by NanoBanana from Google (Image Result Sets Dataset) | Setting | Value | | --------------------------- | --------- | | Seed | - | | Clip Skip | - | | Text Encoder LR | 0.00001 | | UNet LR | 0.00005 | | LR Scheduler | constant | | Optimizer | AdamW8bit | | Network Dimension | 64 | | Network Alpha | 32 | | Gradient Accumulation Steps | - | | Setting | Value | | --------------- | ----- | | Shuffle Caption | - | | Keep N Tokens | - | | Setting | Value | | ------------------------- | ----- | | Noise Offset | 0.03 | | Multires Noise Discount | 0.1 | | Multires Noise Iterations | 10 | | Conv Dimension | - | | Conv Alpha | - | | Batch Size | - | | Steps | 4100 | | Sampler | euler | You should use `[photo content]` to trigger the image generation. You should use `restore and enhance the image by repairing any damage` to trigger the image generation. You should use `scratches` to trigger the image generation. You should use `or fading. Colorize the photo naturally while preserving authentic textures and details` to trigger the image generation. You should use `maintaining a realistic and historically accurate look.` to trigger the image generation.

> Speech-Emotion-Classification is a fine-tuned version of `facebook/wav2vec2-base-960h` for multi-class audio classification, specifically trained to detect emotions in speech. This model utilizes the `Wav2Vec2ForSequenceClassification` architecture to accurately classify speaker emotions from audio signals. > \[!note] > Wav2Vec2: Self-Supervised Learning for Speech Recognition > https://arxiv.org/pdf/2006.11477 Speech Emotion Analytics – Analyze speaker emotions in call centers, interviews, or therapeutic sessions. Conversational AI Personalization – Adjust voice assistant responses based on detected emotion. Mental Health Monitoring – Support emotion recognition in voice-based wellness or teletherapy apps. Voice Dataset Curation – Tag or filter speech datasets by emotion for research or model training. Media Annotation – Automatically annotate podcasts, audiobooks, or videos with speaker emotion metadata.

> ZeroXClem/Qwen-4B-Valiant-Polaris is a thoughtfully merged 4B-parameter language model built upon Qwen3-4B, combining the structured reasoning of Polaris, the creative and expressive capabilities of Dot-Goat and RP-V3, and the scientific depth of ShiningValiant3, resulting in a lightweight yet powerful architecture designed for advanced reasoning, rich roleplay, scientific and analytical tasks, and seamless agentic workflows; with robust support for long contexts, multilingual reasoning, and tool integration, it is ideal for conversational agents, tutoring, problem solving, creative writing, and autonomous agent applications. | File name | Size | Quant type | |-----------|------|------------| | Qwen3-4B-Valiant-Polaris.BF16.gguf | 8.05 GB | BF16 | | Qwen3-4B-Valiant-Polaris.F16.gguf | 8.05 GB | F16 | | Qwen3-4B-Valiant-Polaris.F32.gguf | 16.1 GB | F32 | | Qwen3-4B-Valiant-Polaris.Q2K.gguf | 1.67 GB | Q2K | | Qwen3-4B-Valiant-Polaris.Q3KL.gguf | 2.24 GB | Q3KL | | Qwen3-4B-Valiant-Polaris.Q3KM.gguf | 2.08 GB | Q3KM | | Qwen3-4B-Valiant-Polaris.Q3KS.gguf | 1.89 GB | Q3KS | | Qwen3-4B-Valiant-Polaris.Q4KM.gguf | 2.5 GB | Q4KM | | Qwen3-4B-Valiant-Polaris.Q4KS.gguf | 2.38 GB | Q4KS | | Qwen3-4B-Valiant-Polaris.Q5KM.gguf | 2.89 GB | Q5KM | | Qwen3-4B-Valiant-Polaris.Q5KS.gguf | 2.82 GB | Q5KS | | Qwen3-4B-Valiant-Polaris.Q6K.gguf | 3.31 GB | Q6K | | Qwen3-4B-Valiant-Polaris.Q80.gguf | 4.28 GB | Q80 | (sorted by size, not necessarily quality. IQ-quants are often preferable over similar sized non-IQ quants) Here is a handy graph by ikawrakow comparing some lower-quality quant types (lower is better):

> Jupyter-agent-qwen3-4b-instruct and jupyter-agent-qwen3-4b-thinking are specialized 4B-parameter models built on Qwen3-4B for agentic reasoning and data science tasks in Jupyter notebooks, supporting both general instruction following and step-by-step, notebook-native logical analysis. The instruct variant excels at delivering fast, efficient responses without generating detailed reasoning traces, while the thinking variant provides comprehensive intermediate computations and analysis, including tool calling and dataset-grounded reasoning on realistic Kaggle workflows, making both models state-of-the-art for code execution, data exploration, and practical problem solving in Python and multi-turn notebook environments. | Model Name | Download Link | |------------------------------------------|-----------------------------------------------------------------------------------------------| | jupyter-agent-qwen3-4b-instruct-GGUF | Link | | jupyter-agent-qwen3-4b-thinking-GGUF | Link | | File Name | Quant Type | File Size | | - | - | - | | jupyter-agent-qwen3-4b-instruct.BF16.gguf | BF16 | 8.05 GB | | jupyter-agent-qwen3-4b-instruct.F16.gguf | F16 | 8.05 GB | | jupyter-agent-qwen3-4b-instruct.F32.gguf | F32 | 16.1 GB | | jupyter-agent-qwen3-4b-instruct.Q2K.gguf | Q2K | 1.67 GB | | jupyter-agent-qwen3-4b-instruct.Q3KL.gguf | Q3KL | 2.24 GB | | jupyter-agent-qwen3-4b-instruct.Q3KM.gguf | Q3KM | 2.08 GB | | jupyter-agent-qwen3-4b-instruct.Q3KS.gguf | Q3KS | 1.89 GB | | jupyter-agent-qwen3-4b-instruct.Q4KM.gguf | Q4KM | 2.5 GB | | jupyter-agent-qwen3-4b-instruct.Q4KS.gguf | Q4KS | 2.38 GB | | jupyter-agent-qwen3-4b-instruct.Q5KM.gguf | Q5KM | 2.89 GB | | jupyter-agent-qwen3-4b-instruct.Q5KS.gguf | Q5KS | 2.82 GB | | jupyter-agent-qwen3-4b-instruct.Q6K.gguf | Q6K | 3.31 GB | | jupyter-agent-qwen3-4b-instruct.Q80.gguf | Q80 | 4.28 GB | | File Name | Quant Type | File Size | | - | - | - | | jupyter-agent-qwen3-4b-thinking.BF16.gguf | BF16 | 8.05 GB | | jupyter-agent-qwen3-4b-thinking.F16.gguf | F16 | 8.05 GB | | jupyter-agent-qwen3-4b-thinking.F32.gguf | F32 | 16.1 GB | | jupyter-agent-qwen3-4b-thinking.Q2K.gguf | Q2K | 1.67 GB | | jupyter-agent-qwen3-4b-thinking.Q3KL.gguf | Q3KL | 2.24 GB | | jupyter-agent-qwen3-4b-thinking.Q3KM.gguf | Q3KM | 2.08 GB | | jupyter-agent-qwen3-4b-thinking.Q3KS.gguf | Q3KS | 1.89 GB | | jupyter-agent-qwen3-4b-thinking.Q4KM.gguf | Q4KM | 2.5 GB | | jupyter-agent-qwen3-4b-thinking.Q4KS.gguf | Q4KS | 2.38 GB | | jupyter-agent-qwen3-4b-thinking.Q5KM.gguf | Q5KM | 2.89 GB | | jupyter-agent-qwen3-4b-thinking.Q5KS.gguf | Q5KS | 2.82 GB | | jupyter-agent-qwen3-4b-thinking.Q6K.gguf | Q6K | 3.31 GB | | jupyter-agent-qwen3-4b-thinking.Q80.gguf | Q80 | 4.28 GB | (sorted by size, not necessarily quality. IQ-quants are often preferable over similar sized non-IQ quants) Here is a handy graph by ikawrakow comparing some lower-quality quant types (lower is better):

> Jan-v1-4B is a 4-billion-parameter language model built on the Qwen3-4B-thinking architecture, meticulously fine-tuned for agentic reasoning, problem-solving, and tool utilization with support for web search tasks and large context lengths up to 256,000 tokens. Achieving 91.1% accuracy on the SimpleQA benchmark, Jan-v1-4B excels at factual question answering and conversation while running efficiently on local hardware for enhanced privacy and offline use, making it a strong choice for advanced Q&A, reasoning, and integration with the Jan desktop application or compatible inference engines. Jan-v1-edge is a lightweight agentic model built for fast, reliable on-device execution. As the second release in the Jan Family, it is distilled from the larger Jan-v1 model, preserving strong reasoning and problem-solving ability in a smaller footprint suitable for resource-constrained environments. Jan-v1-edge was developed through a two-phase post-training process. The first phase, Supervised Fine-Tuning (SFT), transferred core capabilities from the Jan-v1 teacher model to the smaller student. The second phase, Reinforcement Learning with Verifiable Rewards (RLVR) —the same method used in Jan-v1 and Lucy—further optimized reasoning efficiency, tool use, and correctness. This staged approach delivers reliable results on complex, interactive workloads. | Model Name | Hugging Face Link | |---------------|-------------------| | Jan-v1-2509-GGUF | 🔗 Link | | Jan-v1-edge-GGUF | 🔗 Link | | Jan-v1-4B-GGUF | 🔗 Link | | File Name | Quant Type | File Size | | - | - | - | | Jan-v1-2509.BF16.gguf | BF16 | 8.05 GB | | Jan-v1-2509.F16.gguf | F16 | 8.05 GB | | Jan-v1-2509.F32.gguf | F32 | 16.1 GB | | Jan-v1-2509.Q2K.gguf | Q2K | 1.67 GB | | Jan-v1-2509.Q3KL.gguf | Q3KL | 2.24 GB | | Jan-v1-2509.Q3KM.gguf | Q3KM | 2.08 GB | | Jan-v1-2509.Q3KS.gguf | Q3KS | 1.89 GB | | Jan-v1-2509.Q4KM.gguf | Q4KM | 2.5 GB | | Jan-v1-2509.Q4KS.gguf | Q4KS | 2.38 GB | | Jan-v1-2509.Q5KM.gguf | Q5KM | 2.89 GB | | Jan-v1-2509.Q5KS.gguf | Q5KS | 2.82 GB | | Jan-v1-2509.Q6K.gguf | Q6K | 3.31 GB | | Jan-v1-2509.Q80.gguf | Q80 | 4.28 GB | | File Name | Quant Type | File Size | | - | - | - | | Jan-v1-edge.BF16.gguf | BF16 | 3.45 GB | | Jan-v1-edge.F16.gguf | F16 | 3.45 GB | | Jan-v1-edge.F32.gguf | F32 | 6.89 GB | | Jan-v1-edge.Q2K.gguf | Q2K | 778 MB | | Jan-v1-edge.Q3KL.gguf | Q3KL | 1 GB | | Jan-v1-edge.Q3KM.gguf | Q3KM | 940 MB | | Jan-v1-edge.Q3KS.gguf | Q3KS | 867 MB | | Jan-v1-edge.Q40.gguf | Q40 | 1.05 GB | | Jan-v1-edge.Q41.gguf | Q41 | 1.14 GB | | Jan-v1-edge.Q4K.gguf | Q4K | 1.11 GB | | Jan-v1-edge.Q4KM.gguf | Q4KM | 1.11 GB | | Jan-v1-edge.Q4KS.gguf | Q4KS | 1.06 GB | | Jan-v1-edge.Q50.gguf | Q50 | 1.23 GB | | Jan-v1-edge.Q51.gguf | Q51 | 1.32 GB | | Jan-v1-edge.Q5K.gguf | Q5K | 1.26 GB | | Jan-v1-edge.Q5KM.gguf | Q5KM | 1.26 GB | | Jan-v1-edge.Q5KS.gguf | Q5KS | 1.23 GB | | Jan-v1-edge.Q6K.gguf | Q6K | 1.42 GB | | Jan-v1-edge.Q80.gguf | Q80 | 1.83 GB | | File Name | Quant Type | File Size | | - | - | - | | Jan-v1-4B.BF16.gguf | BF16 | 8.05 GB | | Jan-v1-4B.F16.gguf | F16 | 8.05 GB | | Jan-v1-4B.F32.gguf | F32 | 16.1 GB | | Jan-v1-4B.Q2K.gguf | Q2K | 1.67 GB | | Jan-v1-4B.Q3KL.gguf | Q3KL | 2.24 GB | | Jan-v1-4B.Q3KM.gguf | Q3KM | 2.08 GB | | Jan-v1-4B.Q3KS.gguf | Q3KS | 1.89 GB | | Jan-v1-4B.Q4KM.gguf | Q4KM | 2.5 GB | | Jan-v1-4B.Q4KS.gguf | Q4KS | 2.38 GB | | Jan-v1-4B.Q5KM.gguf | Q5KM | 2.89 GB | | Jan-v1-4B.Q5KS.gguf | Q5KS | 2.82 GB | | Jan-v1-4B.Q6K.gguf | Q6K | 3.31 GB | | Jan-v1-4B.Q80.gguf | Q80 | 4.28 GB | (sorted by size, not necessarily quality. IQ-quants are often preferable over similar sized non-IQ quants) Here is a handy graph by ikawrakow comparing some lower-quality quant types (lower is better):

> Qwen3MedicalGRPO is a specialized medical language model fine-tuned from the Qwen3 base using Supervised Fine-Tuning (SFT) and enhanced with Group Relative Policy Optimization (GRPO) to deliver advanced performance in clinical case analysis, differential diagnosis, and medical reasoning tasks. The model is designed to provide both detailed, step-by-step reasoning (chain-of-thought) and clear, structured final answers, enabling greater transparency and reliability for healthcare professionals and research applications. By separating its internal analysis from synthesized conclusions, Qwen3MedicalGRPO allows users to trace the logic behind clinical recommendations, optimizing accuracy and trustworthiness in complex medical scenarios. | File Name | Quant Type | File Size | | - | - | - | | Qwen3-Medical-GRPO.BF16.gguf | BF16 | 8.05 GB | | Qwen3-Medical-GRPO.F16.gguf | F16 | 8.05 GB | | Qwen3-Medical-GRPO.F32.gguf | F32 | 16.1 GB | | Qwen3-Medical-GRPO.Q2K.gguf | Q2K | 1.67 GB | | Qwen3-Medical-GRPO.Q3KL.gguf | Q3KL | 2.24 GB | | Qwen3-Medical-GRPO.Q3KM.gguf | Q3KM | 2.08 GB | | Qwen3-Medical-GRPO.Q3KS.gguf | Q3KS | 1.89 GB | | Qwen3-Medical-GRPO.Q4KM.gguf | Q4KM | 2.5 GB | | Qwen3-Medical-GRPO.Q4KS.gguf | Q4KS | 2.38 GB | | Qwen3-Medical-GRPO.Q5KM.gguf | Q5KM | 2.89 GB | | Qwen3-Medical-GRPO.Q5KS.gguf | Q5KS | 2.82 GB | | Qwen3-Medical-GRPO.Q6K.gguf | Q6K | 3.31 GB | | Qwen3-Medical-GRPO.Q80.gguf | Q80 | 4.28 GB | (sorted by size, not necessarily quality. IQ-quants are often preferable over similar sized non-IQ quants) Here is a handy graph by ikawrakow comparing some lower-quality quant types (lower is better):

> Qwen3-Reranker-0.6B-seq-cls is a sequence classification adaptation of the Qwen3-Reranker-0.6B model, designed for advanced text reranking and classification tasks across over 100 languages, including code and multilingual retrieval scenarios. Built on the Qwen3 series, this 0.6B parameter model offers a 32k context window and supports instruction-aware customization for specific tasks, typically improving performance by 1% to 5% when using tailored English instructions. It inherits the robust reasoning, long-text understanding, and versatility of its parent models, excelling in text retrieval, code retrieval, clustering, classification, and bitext mining, and ranks among the top models in benchmarks such as the MTEB multilingual leaderboard. | File name | Size | Quant Type | |-----------|------|------------| | Qwen3-4B-abliterated.F32.gguf | 16.1 GB | F32 | | Qwen3-4B-abliterated.BF16.gguf | 8.05 GB | BF16 | | Qwen3-4B-abliterated.F16.gguf | 8.05 GB | F16 | | Qwen3-4B-abliterated.Q80.gguf | 4.28 GB | Q80 | | Qwen3-4B-abliterated.Q6K.gguf | 3.31 GB | Q6K | | Qwen3-4B-abliterated.Q5KM.gguf | 2.89 GB | Q5KM | | Qwen3-4B-abliterated.Q5KS.gguf | 2.82 GB | Q5KS | | Qwen3-4B-abliterated.Q4KM.gguf | 2.5 GB | Q4KM | | Qwen3-4B-abliterated.Q4KS.gguf | 2.38 GB | Q4KS | | Qwen3-4B-abliterated.Q3KL.gguf | 2.24 GB | Q3KL | | Qwen3-4B-abliterated.Q3KM.gguf | 2.08 GB | Q3KM | | Qwen3-4B-abliterated.Q3KS.gguf | 1.89 GB | Q3KS | | Qwen3-4B-abliterated.Q2K.gguf | 1.67 GB | Q2K | (sorted by size, not necessarily quality. IQ-quants are often preferable over similar sized non-IQ quants) Here is a handy graph by ikawrakow comparing some lower-quality quant types (lower is better):

> VyvoTTS-v0-Qwen3-0.6B is an English Text-to-Speech (TTS) model built on the Qwen3-0.6B architecture and trained using a 10,000-hour dataset to produce natural-sounding speech. With approximately 810 million parameters and licensed under MIT, the model offers flexible usage as a pretrained base for further development, especially recommended to be enhanced by leveraging the Emilia dataset and fine-tuning for single-speaker scenarios. Users can integrate VyvoTTS with the unsloth and SNAC frameworks for speech generation, and the model supports sequence lengths up to 8,192 tokens. Although it currently exhibits a high Word Error Rate (WER), its open-source nature and compatibility with popular Python libraries make it an accessible starting point for advanced speech synthesis projects. | File Name | Size | Quant Type | |-----------|------|------------| | VyvoTTS-v0-Qwen3-0.6B.BF16.gguf | 1.26 GB | BF16 | | VyvoTTS-v0-Qwen3-0.6B.F16.gguf | 1.26 GB | F16 | | VyvoTTS-v0-Qwen3-0.6B.F32.gguf | 2.51 GB | F32 | | VyvoTTS-v0-Qwen3-0.6B.Q2K.gguf | 321 MB | Q2K | | VyvoTTS-v0-Qwen3-0.6B.Q3KL.gguf | 393 MB | Q3KL | | VyvoTTS-v0-Qwen3-0.6B.Q3KM.gguf | 372 MB | Q3KM | | VyvoTTS-v0-Qwen3-0.6B.Q3KS.gguf | 348 MB | Q3KS | | VyvoTTS-v0-Qwen3-0.6B.Q40.gguf | 406 MB | Q40 | | VyvoTTS-v0-Qwen3-0.6B.Q41.gguf | 434 MB | Q41 | | VyvoTTS-v0-Qwen3-0.6B.Q4K.gguf | 421 MB | Q4K | | VyvoTTS-v0-Qwen3-0.6B.Q4KM.gguf | 421 MB | Q4KM | | VyvoTTS-v0-Qwen3-0.6B.Q4KS.gguf | 408 MB | Q4KS | | VyvoTTS-v0-Qwen3-0.6B.Q50.gguf | 461 MB | Q50 | | VyvoTTS-v0-Qwen3-0.6B.Q51.gguf | 489 MB | Q51 | | VyvoTTS-v0-Qwen3-0.6B.Q5K.gguf | 469 MB | Q5K | | VyvoTTS-v0-Qwen3-0.6B.Q5KM.gguf | 469 MB | Q5KM | | VyvoTTS-v0-Qwen3-0.6B.Q5KS.gguf | 461 MB | Q5KS | | VyvoTTS-v0-Qwen3-0.6B.Q6K.gguf | 520 MB | Q6K | | VyvoTTS-v0-Qwen3-0.6B.Q80.gguf | 671 MB | Q80 | (sorted by size, not necessarily quality. IQ-quants are often preferable over similar sized non-IQ quants) Here is a handy graph by ikawrakow comparing some lower-quality quant types (lower is better):

The Kontext-Watermark-Remover is an adapter for black-forest-lab's FLUX.1-Kontext-dev, designed to precisely remove watermarks and textual content from images while maintaining the original image quality and context. The model was trained on 150 image pairs (75 start images and 75 end images) to ensure accurate and artifact-free watermark removal. > [!note] [photo content], remove any watermark text or logos from the image while preserving the background, texture, lighting, and overall realism. Ensure the edited areas blend seamlessly with surrounding details, leaving no visible traces of watermark removal. | Setting | Value | | ------------------------ | ------------------------ | | Module Type | Adapter | | Base Model | FLUX.1 Kontext Dev - fp8 | | Trigger Words | [photo content], remove any watermark text or logos from the image while preserving the background, texture, lighting, and overall realism. Ensure the edited areas blend seamlessly with surrounding details, leaving no visible traces of watermark removal. | | Image Processing Repeats | 50 | | Epochs | 25 | | Save Every N Epochs | 1 | Labeling: florence-community/Florence-2-large-ft (natural language & English) Total Images Used for Training : 150 Image Pairs (75 Start, 75 End) | Setting | Value | | --------------------------- | --------- | | Seed | - | | Clip Skip | - | | Text Encoder LR | 0.00001 | | UNet LR | 0.00005 | | LR Scheduler | constant | | Optimizer | AdamW8bit | | Network Dimension | 64 | | Network Alpha | 32 | | Gradient Accumulation Steps | - | | Setting | Value | | --------------- | ----- | | Shuffle Caption | - | | Keep N Tokens | - | | Setting | Value | | ------------------------- | ----- | | Noise Offset | 0.03 | | Multires Noise Discount | 0.1 | | Multires Noise Iterations | 10 | | Conv Dimension | - | | Conv Alpha | - | | Batch Size | - | | Steps | 2900 & 400(warm up) | | Sampler | euler | You should use `[photo content]` to trigger the image generation. You should use `remove any watermark text or logos from the image while preserving the background` to trigger the image generation. You should use `texture` to trigger the image generation. You should use `lighting` to trigger the image generation. You should use `and overall realism. Ensure the edited areas blend seamlessly with surrounding details` to trigger the image generation. You should use `leaving no visible traces of watermark removal.` to trigger the image generation.

> Octans-Qwen3-UI-Code-4B is an optimized successor of Muscae-Qwen3-UI-Code-4B, fine-tuned for enhanced UI reasoning precision, layout structuring, and frontend code synthesis. Built upon Qwen3-4B and refined through Abliterated Reasoning Optimization, it delivers balanced, structured, and production-grade UI code outputs for experimental and research use. Ideal for frontend developers, UI engineers, and design system researchers exploring next-generation code synthesis. | File Name | Quant Type | File Size | | - | - | - | | Octans-Qwen3-UI-Code-4B.BF16.gguf | BF16 | 8.05 GB | | Octans-Qwen3-UI-Code-4B.F16.gguf | F16 | 8.05 GB | | Octans-Qwen3-UI-Code-4B.F32.gguf | F32 | 16.1 GB | | Octans-Qwen3-UI-Code-4B.Q2K.gguf | Q2K | 1.67 GB | | Octans-Qwen3-UI-Code-4B.Q3KL.gguf | Q3KL | 2.24 GB | | Octans-Qwen3-UI-Code-4B.Q3KM.gguf | Q3KM | 2.08 GB | | Octans-Qwen3-UI-Code-4B.Q3KS.gguf | Q3KS | 1.89 GB | | Octans-Qwen3-UI-Code-4B.Q4KM.gguf | Q4KM | 2.5 GB | | Octans-Qwen3-UI-Code-4B.Q4KS.gguf | Q4KS | 2.38 GB | | Octans-Qwen3-UI-Code-4B.Q5KM.gguf | Q5KM | 2.89 GB | | Octans-Qwen3-UI-Code-4B.Q5KS.gguf | Q5KS | 2.82 GB | | Octans-Qwen3-UI-Code-4B.Q6K.gguf | Q6K | 3.31 GB | | Octans-Qwen3-UI-Code-4B.Q80.gguf | Q80 | 4.28 GB | (sorted by size, not necessarily quality. IQ-quants are often preferable over similar sized non-IQ quants) Here is a handy graph by ikawrakow comparing some lower-quality quant types (lower is better):

Deepfake-Detection-Exp-02-21 is a minimalist, high-quality dataset trained on a ViT-based model for image classification, distinguishing between deepfake and real images. The model is based on Google's `google/vit-base-patch16-224-in21k`. Limitations 1. Generalization Issues – The model may not perform well on deepfake images generated by unseen or novel deepfake techniques. 2. Dataset Bias – The training data might not cover all variations of real and fake images, leading to biased predictions. 3. Resolution Constraints – Since the model is based on `vit-base-patch16-224-in21k`, it is optimized for 224x224 image resolution, which may limit its effectiveness on high-resolution images. 4. Adversarial Vulnerabilities – The model may be susceptible to adversarial attacks designed to fool vision transformers. 5. False Positives & False Negatives – The model may occasionally misclassify real images as deepfake and vice versa, requiring human validation in critical applications. Intended Use 1. Deepfake Detection – Designed for identifying deepfake images in media, social platforms, and forensic analysis. 2. Research & Development – Useful for researchers studying deepfake detection and improving ViT-based classification models. 3. Content Moderation – Can be integrated into platforms to detect and flag manipulated images. 4. Security & Forensics – Assists in cybersecurity applications where verifying the authenticity of images is crucial. 5. Educational Purposes – Can be used in training AI practitioners and students in the field of computer vision and deepfake detection.

PhotoCleanser-i2i is an adapter for black-forest-lab's FLUX.1-Kontext-dev. It is an experimental LoRA designed for removing specified object(s) while preserving the remaining content in the image. The model was trained on 36 image pairs (18 start images, 18 end images). Synthetic result nodes were generated using NanoBanana from Google and labeled with DeepCaption-VLA-7B. The adapter is triggered with the following prompt: > [!note] [photo content], remove the specified object(s) from the image while preserving the background and remaining elements, maintaining realism and original details. > [photo content], remove all humans from the image while preserving the background and remaining elements, maintaining realism and original details. > [photo content], remove the ball from the image while preserving the background and remaining elements, maintaining realism and original details. > [photo content], remove the cat from the image while preserving the background and remaining elements, maintaining realism and original details. Sample Inference Comparing the Base Model with the Adapter > Note: In over 25 inferences across various scale settings, the base model struggles to properly reconstruct the basketball net after the ball is removed. | FLUX.1-Kontext-dev | PhotoCleanser-i2i | |-------------------|-----------------| | | | > No desired effect was achieved with the base model after testing many settings, but PhotoCleanser-i2i performed its best to remove the human character from the image. | FLUX.1-Kontext-dev | PhotoCleanser-i2i | |-------------------|-----------------| | | | > No desired effect was achieved with the base model after testing many settings, but PhotoCleanser-i2i performed its best to remove the human/animal character from the image. | FLUX.1-Kontext-dev | PhotoCleanser-i2i | |-------------------|-----------------| | | | | Setting | Value | | ------------------------ | ------------------------ | | Module Type | Adapter | | Base Model | FLUX.1 Kontext Dev - fp8 | | Trigger Words | [photo content], remove the specified object(s) from the image while preserving the background and remaining elements, maintaining realism and original details. | | Image Processing Repeats | 50 | | Epochs | 32 | | Save Every N Epochs | 1 | Labeling: DeepCaption-VLA-7B(natural language & English) Total Images Used for Training : 36 Image Pairs (18 Start, 18 End) Synthetic Result Node generated by NanoBanana from Google (Image Result Sets Dataset) | Setting | Value | | --------------------------- | --------- | | Seed | - | | Clip Skip | - | | Text Encoder LR | 0.00001 | | UNet LR | 0.00005 | | LR Scheduler | constant | | Optimizer | AdamW8bit | | Network Dimension | 64 | | Network Alpha | 32 | | Gradient Accumulation Steps | - | | Setting | Value | | --------------- | ----- | | Shuffle Caption | - | | Keep N Tokens | - | | Setting | Value | | ------------------------- | ----- | | Noise Offset | 0.03 | | Multires Noise Discount | 0.1 | | Multires Noise Iterations | 10 | | Conv Dimension | - | | Conv Alpha | - | | Batch Size | - | | Steps | 4370 (Low(700)) | | Sampler | euler | You should use `[photo content]` to trigger the image generation. You should use `remove the specified object(s) from the image while preserving the background and remaining elements` to trigger the image generation. You should use `maintaining realism and original details.` to trigger the image generation.

> Gliese-4B-OSS-0410 is a reasoning-focused model fine-tuned on Qwen-4B for enhanced reasoning and polished token probability distributions, delivering balanced multilingual generation across mathematics and general-purpose reasoning tasks. The model is fine-tuned on curated GPT-OSS synthetic dataset entries, improving its ability to handle structured reasoning, probabilistic inference, and multilingual tasks with precision. | File Name | Quant Type | File Size | | - | - | - | | Gliese-4B-OSS-0410.BF16.gguf | BF16 | 8.05 GB | | Gliese-4B-OSS-0410.F16.gguf | F16 | 8.05 GB | | Gliese-4B-OSS-0410.F32.gguf | F32 | 16.1 GB | | Gliese-4B-OSS-0410.Q2K.gguf | Q2K | 1.67 GB | | Gliese-4B-OSS-0410.Q3KL.gguf | Q3KL | 2.24 GB | | Gliese-4B-OSS-0410.Q3KM.gguf | Q3KM | 2.08 GB | | Gliese-4B-OSS-0410.Q3KS.gguf | Q3KS | 1.89 GB | | Gliese-4B-OSS-0410.Q4KM.gguf | Q4KM | 2.5 GB | | Gliese-4B-OSS-0410.Q4KS.gguf | Q4KS | 2.38 GB | | Gliese-4B-OSS-0410.Q5KM.gguf | Q5KM | 2.89 GB | | Gliese-4B-OSS-0410.Q5KS.gguf | Q5KS | 2.82 GB | | Gliese-4B-OSS-0410.Q6K.gguf | Q6K | 3.31 GB | | Gliese-4B-OSS-0410.Q80.gguf | Q80 | 4.28 GB | (sorted by size, not necessarily quality. IQ-quants are often preferable over similar sized non-IQ quants) Here is a handy graph by ikawrakow comparing some lower-quality quant types (lower is better):

> MegaScience-Qwen models are a series of large language models based on the Qwen3 and Qwen2.5 architectures, meticulously fine-tuned on the MegaScience dataset to advance scientific reasoning in AI. This dataset blends over 1.25 million high-quality, university-level scientific questions and answers sourced from open textbooks and diverse scientific benchmarks, covering seven scientific disciplines. The MegaScience-Qwen lineup includes variants from smaller Qwen2.5-1.5B up to Qwen3-30B, with models such as Qwen3-4B-MegaScience, Qwen3-8B-MegaScience, and Qwen3-14B-MegaScience, each showing pronounced gains over their official instruction-tuned counterparts—especially as model scale increases. These models demonstrate state-of-the-art or leading performance on scientific reasoning, general knowledge, and mathematical benchmarks, achieving not only higher accuracy but also more concise and efficient responses. The MegaScience project also provides a rigorous evaluation system, an open-source curation pipeline, and all model checkpoints, empowering further research and application in scientific AI reasoning and education. | Model Name | GGUF Repository Link | |--------------------------------|----------------------------------------------------------------------------------------| | Qwen3-8B-MegaScience-GGUF | Hugging Face ↗| | Qwen3-4B-MegaScience-GGUF | Hugging Face ↗ | | Qwen3-1.7B-MegaScience-GGUF | Hugging Face ↗ | | Qwen2.5-3B-MegaScience-GGUF | Hugging Face ↗ | | Qwen2.5-1.5B-MegaScience-GGUF | Hugging Face ↗ | | Qwen2.5-7B-MegaScience-GGUF | Hugging Face ↗ | | File Name | Quant Type | File Size | | - | - | - | | Qwen3-8B-MegaScience.BF16.gguf | BF16 | 16.4 GB | | Qwen3-8B-MegaScience.F16.gguf | F16 | 16.4 GB | | Qwen3-8B-MegaScience.Q80.gguf | Q80 | 8.71 GB | | File Name | Size | Quant Type | |-----------|------|------------| | Qwen3-4B-MegaScience.BF16.gguf | 8.05 GB | BF16 | | Qwen3-4B-MegaScience.F16.gguf | 8.05 GB | F16 | | Qwen3-4B-MegaScience.F32.gguf | 16.1 GB | F32 | | Qwen3-4B-MegaScience.Q2K.gguf | 1.67 GB | Q2K | | Qwen3-4B-MegaScience.Q3KL.gguf | 2.24 GB | Q3KL | | Qwen3-4B-MegaScience.Q3KM.gguf | 2.08 GB | Q3KM | | Qwen3-4B-MegaScience.Q3KS.gguf | 1.89 GB | Q3KS | | Qwen3-4B-MegaScience.Q4KM.gguf | 2.5 GB | Q4KM | | Qwen3-4B-MegaScience.Q4KS.gguf | 2.38 GB | Q4KS | | Qwen3-4B-MegaScience.Q5KM.gguf | 2.89 GB | Q5KM | | Qwen3-4B-MegaScience.Q5KS.gguf | 2.82 GB | Q5KS | | Qwen3-4B-MegaScience.Q6K.gguf | 3.31 GB | Q6K | | Qwen3-4B-MegaScience.Q80.gguf | 4.28 GB | Q80 | | File Name | Size | Quant Type | |-----------|------|------------| | Qwen3-1.7B-MegaScience.BF16.gguf | 3.45 GB | BF16 | | Qwen3-1.7B-MegaScience.F16.gguf | 3.45 GB | F16 | | Qwen3-1.7B-MegaScience.F32.gguf | 6.89 GB | F32 | | Qwen3-1.7B-MegaScience.Q2K.gguf | 778 MB | Q2K | | Qwen3-1.7B-MegaScience.Q3KL.gguf | 1 GB | Q3KL | | Qwen3-1.7B-MegaScience.Q3KM.gguf | 940 MB | Q3KM | | Qwen3-1.7B-MegaScience.Q3KS.gguf | 867 MB | Q3KS | | Qwen3-1.7B-MegaScience.Q4KM.gguf | 1.11 GB | Q4KM | | Qwen3-1.7B-MegaScience.Q4KS.gguf | 1.06 GB | Q4KS | | Qwen3-1.7B-MegaScience.Q5KM.gguf | 1.26 GB | Q5KM | | Qwen3-1.7B-MegaScience.Q5KS.gguf | 1.23 GB | Q5KS | | Qwen3-1.7B-MegaScience.Q6K.gguf | 1.42 GB | Q6K | | Qwen3-1.7B-MegaScience.Q80.gguf | 1.83 GB | Q80 | | File Name | Size | Quant Type | |-----------|------|------------| | Qwen2.5-3B-MegaScience.BF16.gguf | 6.18 GB | BF16 | | Qwen2.5-3B-MegaScience.F16.gguf | 6.18 GB | F16 | | Qwen2.5-3B-MegaScience.F32.gguf | 12.3 GB | F32 | | Qwen2.5-3B-MegaScience.Q2K.gguf | 1.27 GB | Q2K | | Qwen2.5-3B-MegaScience.Q3KL.gguf | 1.71 GB | Q3KL | | Qwen2.5-3B-MegaScience.Q3KM.gguf | 1.59 GB | Q3KM | | Qwen2.5-3B-MegaScience.Q3KS.gguf | 1.45 GB | Q3KS | | Qwen2.5-3B-MegaScience.Q4KM.gguf | 1.93 GB | Q4KM | | Qwen2.5-3B-MegaScience.Q4KS.gguf | 1.83 GB | Q4KS | | Qwen2.5-3B-MegaScience.Q5KM.gguf | 2.22 GB | Q5KM | | Qwen2.5-3B-MegaScience.Q5KS.gguf | 2.17 GB | Q5KS | | Qwen2.5-3B-MegaScience.Q6K.gguf | 2.54 GB | Q6K | | Qwen2.5-3B-MegaScience.Q80.gguf | 3.29 GB | Q80 | | File Name | Size | Quant Type | |-----------|------|------------| | Qwen2.5-1.5B-MegaScience.BF16.gguf | 3.09 GB | BF16 | | Qwen2.5-1.5B-MegaScience.F16.gguf | 3.09 GB | F16 | | Qwen2.5-1.5B-MegaScience.F32.gguf | 6.18 GB | F32 | | Qwen2.5-1.5B-MegaScience.Q2K.gguf | 676 MB | Q2K | | Qwen2.5-1.5B-MegaScience.Q3KL.gguf | 880 MB | Q3KL | | Qwen2.5-1.5B-MegaScience.Q3KM.gguf | 824 MB | Q3KM | | Qwen2.5-1.5B-MegaScience.Q3KS.gguf | 761 MB | Q3KS | | Qwen2.5-1.5B-MegaScience.Q4KM.gguf | 986 MB | Q4KM | | Qwen2.5-1.5B-MegaScience.Q4KS.gguf | 940 MB | Q4KS | | Qwen2.5-1.5B-MegaScience.Q5KM.gguf | 1.13 GB | Q5KM | | Qwen2.5-1.5B-MegaScience.Q5KS.gguf | 1.1 GB | Q5KS | | Qwen2.5-1.5B-MegaScience.Q6K.gguf | 1.27 GB | Q6K | | Qwen2.5-1.5B-MegaScience.Q80.gguf | 1.65 GB | Q80 | | File Name | Quant Type | File Size | | - | - | - | | Qwen2.5-7B-MegaScience.BF16.gguf | BF16 | 15.2 GB | | Qwen2.5-7B-MegaScience.F16.gguf | F16 | 15.2 GB | | Qwen2.5-7B-MegaScience.F32.gguf | F32 | 30.5 GB | (sorted by size, not necessarily quality. IQ-quants are often preferable over similar sized non-IQ quants) Here is a handy graph by ikawrakow comparing some lower-quality quant types (lower is better):

> Qwen3-4B-abliterated is an experimental, uncensored version of the Qwen/Qwen3-4B language model that explores how refusals and latent fine-tuning work in large language models using a novel "abliteration" technique, which subtracts a computed refusal direction from hidden module states (such as oproj) to minimize refusals without degrading output quality. The process involves comparing residual streams between harmful and harmless prompts, orthogonalizing hidden states with weight factors distributed across layers, and iterative or accumulated orthogonalization methods for efficiency. | File name | Size | Quant Type | |-----------|------|------------| | Qwen3-4B-abliterated.F32.gguf | 16.1 GB | F32 | | Qwen3-4B-abliterated.BF16.gguf | 8.05 GB | BF16 | | Qwen3-4B-abliterated.F16.gguf | 8.05 GB | F16 | | Qwen3-4B-abliterated.Q80.gguf | 4.28 GB | Q80 | | Qwen3-4B-abliterated.Q6K.gguf | 3.31 GB | Q6K | | Qwen3-4B-abliterated.Q5KM.gguf | 2.89 GB | Q5KM | | Qwen3-4B-abliterated.Q5KS.gguf | 2.82 GB | Q5KS | | Qwen3-4B-abliterated.Q4KM.gguf | 2.5 GB | Q4KM | | Qwen3-4B-abliterated.Q4KS.gguf | 2.38 GB | Q4KS | | Qwen3-4B-abliterated.Q3KL.gguf | 2.24 GB | Q3KL | | Qwen3-4B-abliterated.Q3KM.gguf | 2.08 GB | Q3KM | | Qwen3-4B-abliterated.Q3KS.gguf | 1.89 GB | Q3KS | | Qwen3-4B-abliterated.Q2K.gguf | 1.67 GB | Q2K | (sorted by size, not necessarily quality. IQ-quants are often preferable over similar sized non-IQ quants) Here is a handy graph by ikawrakow comparing some lower-quality quant types (lower is better):

> `open-age-detection` is a vision-language encoder model fine-tuned from `google/siglip2-base-patch16-512` for multi-class image classification. It is trained to classify the estimated age group of a person from an image. The model uses the `SiglipForImageClassification` architecture. > \[!note] > SigLIP 2: Multilingual Vision-Language Encoders with Improved Semantic Understanding, Localization, and Dense Features > https://arxiv.org/pdf/2502.14786 Demographic Analysis – Estimate age groups for statistical or analytical applications. Smart Personalization – Age-based content or product recommendation. Access Control – Assist systems requiring age verification. Social Research – Study age-related trends in image datasets. Surveillance and Security – Profile age ranges in monitored environments.

> Regulus-Qwen3-R1-Llama-Distill-1.7B is a distilled reasoning model fine-tuned on Qwen/Qwen3-1.7B using Magpie-Align/Magpie-Reasoning-V2-250K-CoT-DeepSeek-R1-Llama-70B. The training leverages distilled traces from DeepSeek-R1-Llama-70B, transferring advanced reasoning patterns into a lightweight 1.7B parameter model. It is specialized for chain-of-thought reasoning across code, math, and science, optimized for efficiency and mid-resource deployment. | File Name | Quant Type | File Size | | - | - | - | | Regulus-Qwen3-R1-Llama-Distill-1.7B.BF16.gguf | BF16 | 3.45 GB | | Regulus-Qwen3-R1-Llama-Distill-1.7B.F16.gguf | F16 | 3.45 GB | | Regulus-Qwen3-R1-Llama-Distill-1.7B.F32.gguf | F32 | 6.89 GB | | Regulus-Qwen3-R1-Llama-Distill-1.7B.Q2K.gguf | Q2K | 778 MB | | Regulus-Qwen3-R1-Llama-Distill-1.7B.Q3KL.gguf | Q3KL | 1 GB | | Regulus-Qwen3-R1-Llama-Distill-1.7B.Q3KM.gguf | Q3KM | 940 MB | | Regulus-Qwen3-R1-Llama-Distill-1.7B.Q3KS.gguf | Q3KS | 867 MB | | Regulus-Qwen3-R1-Llama-Distill-1.7B.Q40.gguf | Q40 | 1.05 GB | | Regulus-Qwen3-R1-Llama-Distill-1.7B.Q41.gguf | Q41 | 1.14 GB | | Regulus-Qwen3-R1-Llama-Distill-1.7B.Q4K.gguf | Q4K | 1.11 GB | | Regulus-Qwen3-R1-Llama-Distill-1.7B.Q4KM.gguf | Q4KM | 1.11 GB | | Regulus-Qwen3-R1-Llama-Distill-1.7B.Q4KS.gguf | Q4KS | 1.06 GB | | Regulus-Qwen3-R1-Llama-Distill-1.7B.Q50.gguf | Q50 | 1.23 GB | | Regulus-Qwen3-R1-Llama-Distill-1.7B.Q51.gguf | Q51 | 1.32 GB | | Regulus-Qwen3-R1-Llama-Distill-1.7B.Q5K.gguf | Q5K | 1.26 GB | | Regulus-Qwen3-R1-Llama-Distill-1.7B.Q5KM.gguf | Q5KM | 1.26 GB | | Regulus-Qwen3-R1-Llama-Distill-1.7B.Q5KS.gguf | Q5KS | 1.23 GB | | Regulus-Qwen3-R1-Llama-Distill-1.7B.Q6K.gguf | Q6K | 1.42 GB | | Regulus-Qwen3-R1-Llama-Distill-1.7B.Q80.gguf | Q80 | 1.83 GB | (sorted by size, not necessarily quality. IQ-quants are often preferable over similar sized non-IQ quants) Here is a handy graph by ikawrakow comparing some lower-quality quant types (lower is better):

> Logics-Qwen3-Math-4B is a reasoning-focused model fine-tuned on Qwen3-4B-Thinking-2507 for mathematical reasoning and logical coding, trained on OpenMathReasoning, OpenCodeReasoning, and Helios-R-6M datasets. It excels in structured mathematical problem solving, algorithmic logic, and probabilistic reasoning, making it ideal for educators, researchers, and developers focused on computational logic and math. 1. Mathematical & Logical Reasoning Fine-tuned for high-precision math reasoning, algorithmic problem-solving, and logical coding tasks. 2. Event-Driven & Probabilistic Modeling Performs probability-based simulations, structured decision-making, and multi-step logical reasoning with strong accuracy. 3. Multilingual Problem Solving Supports math and logic tasks across multiple languages, suitable for global research and education workflows. 4. Hybrid Symbolic-Algorithmic Thinking Combines structured logic, symbolic computation, and probabilistic inference to handle uncertainty-driven problems efficiently. 5. Structured Output Mastery Generates outputs in LaTeX, Markdown, JSON, CSV, and YAML, enabling smooth integration into technical and research workflows. 6. Optimized 4B Parameter Footprint Deployable on mid-range GPUs, offline clusters, and edge devices, maintaining high reasoning quality while being resource-efficient. High-precision mathematical reasoning and problem-solving Algorithmic logic, structured coding tasks, and probability analysis Educational and research-focused workflows Deployment on mid-resource environments with efficient reasoning Structured data and technical content generation Focused on math and logic—less suited for creative writing or casual conversation Very complex multi-hop reasoning may challenge the 4B parameter capacity Prioritizes structured reasoning over conversational tone Outputs may be inconsistent for extremely long or cross-domain multi-document contexts

> Qwen3-4B-SafeRL is a safety-aligned version of the Qwen3-4B model, trained using Reinforcement Learning (RL) with a reward signal from Qwen3Guard-Gen to boost robustness against harmful or adversarial prompts. This safety alignment process optimizes the model with a hybrid reward function that simultaneously focuses on three objectives: maximizing safety (penalizing unsafe content as detected by Qwen3Guard-Gen-4B), maximizing helpfulness (rewarding genuinely helpful responses based on the WorldPM-Helpsteer2 model), and minimizing unnecessary refusals (penalizing unnecessary refusals according to Qwen3Guard-Gen-4B). | File Name | Quant Type | File Size | | - | - | - | | Qwen3-4B-SafeRL.BF16.gguf | BF16 | 8.05 GB | | Qwen3-4B-SafeRL.F16.gguf | F16 | 8.05 GB | | Qwen3-4B-SafeRL.F32.gguf | F32 | 16.1 GB | | Qwen3-4B-SafeRL.Q2K.gguf | Q2K | 1.67 GB | | Qwen3-4B-SafeRL.Q3KL.gguf | Q3KL | 2.24 GB | | Qwen3-4B-SafeRL.Q3KM.gguf | Q3KM | 2.08 GB | | Qwen3-4B-SafeRL.Q3KS.gguf | Q3KS | 1.89 GB | | Qwen3-4B-SafeRL.Q4KM.gguf | Q4KM | 2.5 GB | | Qwen3-4B-SafeRL.Q4KS.gguf | Q4KS | 2.38 GB | | Qwen3-4B-SafeRL.Q5KM.gguf | Q5KM | 2.89 GB | | Qwen3-4B-SafeRL.Q5KS.gguf | Q5KS | 2.82 GB | | Qwen3-4B-SafeRL.Q6K.gguf | Q6K | 3.31 GB | | Qwen3-4B-SafeRL.Q80.gguf | Q80 | 4.28 GB | (sorted by size, not necessarily quality. IQ-quants are often preferable over similar sized non-IQ quants) Here is a handy graph by ikawrakow comparing some lower-quality quant types (lower is better):

> Kepler-Qwen3-4B-Super-Thinking is a reasoning-focused model fine-tuned on Qwen for Abliterated Reasoning and polished token probabilities, enhancing balanced multilingual generation across mathematics and general-purpose reasoning. It specializes in event-driven logic, structured analysis, and precise probabilistic modeling—making it an ideal tool for researchers, educators, and developers working with uncertainty and structured reasoning. | File Name | Quant Type | File Size | | - | - | - | | Kepler-Qwen3-4B-Super-Thinking.BF16.gguf | BF16 | 8.05 GB | | Kepler-Qwen3-4B-Super-Thinking.F16.gguf | F16 | 8.05 GB | | Kepler-Qwen3-4B-Super-Thinking.F32.gguf | F32 | 16.1 GB | | Kepler-Qwen3-4B-Super-Thinking.Q2K.gguf | Q2K | 1.67 GB | | Kepler-Qwen3-4B-Super-Thinking.Q3KL.gguf | Q3KL | 2.24 GB | | Kepler-Qwen3-4B-Super-Thinking.Q3KM.gguf | Q3KM | 2.08 GB | | Kepler-Qwen3-4B-Super-Thinking.Q3KS.gguf | Q3KS | 1.89 GB | | Kepler-Qwen3-4B-Super-Thinking.Q4KM.gguf | Q4KM | 2.5 GB | | Kepler-Qwen3-4B-Super-Thinking.Q4KS.gguf | Q4KS | 2.38 GB | | Kepler-Qwen3-4B-Super-Thinking.Q5KM.gguf | Q5KM | 2.89 GB | | Kepler-Qwen3-4B-Super-Thinking.Q5KS.gguf | Q5KS | 2.82 GB | | Kepler-Qwen3-4B-Super-Thinking.Q6K.gguf | Q6K | 3.31 GB | | Kepler-Qwen3-4B-Super-Thinking.Q80.gguf | Q80 | 4.28 GB | (sorted by size, not necessarily quality. IQ-quants are often preferable over similar sized non-IQ quants) Here is a handy graph by ikawrakow comparing some lower-quality quant types (lower is better):

> Dorado-WebSurfTool-ext is a function-calling and agentic reasoning model fine-tuned from Qwen3-4B, designed for web search orchestration, tool-augmented reasoning, and dynamic problem-solving. It excels at agentic decision-making, tool selection, and structured execution flow, making it ideal for retrieval-augmented generation (RAG), function calling, and tool-based query resolution. | File Name | Quant Type | File Size | | - | - | - | | Dorado-WebSurfTool-ext.BF16.gguf | BF16 | 8.05 GB | | Dorado-WebSurfTool-ext.F16.gguf | F16 | 8.05 GB | | Dorado-WebSurfTool-ext.F32.gguf | F32 | 16.1 GB | | Dorado-WebSurfTool-ext.Q2K.gguf | Q2K | 1.67 GB | | Dorado-WebSurfTool-ext.Q3KL.gguf | Q3KL | 2.24 GB | | Dorado-WebSurfTool-ext.Q3KM.gguf | Q3KM | 2.08 GB | | Dorado-WebSurfTool-ext.Q3KS.gguf | Q3KS | 1.89 GB | | Dorado-WebSurfTool-ext.Q4KM.gguf | Q4KM | 2.5 GB | | Dorado-WebSurfTool-ext.Q4KS.gguf | Q4KS | 2.38 GB | | Dorado-WebSurfTool-ext.Q5KM.gguf | Q5KM | 2.89 GB | | Dorado-WebSurfTool-ext.Q5KS.gguf | Q5KS | 2.82 GB | | Dorado-WebSurfTool-ext.Q6K.gguf | Q6K | 3.31 GB | | Dorado-WebSurfTool-ext.Q80.gguf | Q80 | 4.28 GB | (sorted by size, not necessarily quality. IQ-quants are often preferable over similar sized non-IQ quants) Here is a handy graph by ikawrakow comparing some lower-quality quant types (lower is better):

> Kepler-186f-Qwen3-Instruct-4B is a reasoning-focused model fine-tuned on Qwen for Abliterated Reasoning and polished token probabilities, enhancing balanced multilingual generation across mathematics and general-purpose reasoning. It specializes in event-driven logic, structured analysis, and precise probabilistic modeling—making it an ideal tool for researchers, educators, and developers working with uncertainty and structured reasoning. | File Name | Quant Type | File Size | | - | - | - | | Kepler-186f-Qwen3-Instruct-4B.BF16.gguf | BF16 | 8.05 GB | | Kepler-186f-Qwen3-Instruct-4B.F16.gguf | F16 | 8.05 GB | | Kepler-186f-Qwen3-Instruct-4B.F32.gguf | F32 | 16.1 GB | | Kepler-186f-Qwen3-Instruct-4B.Q2K.gguf | Q2K | 1.67 GB | | Kepler-186f-Qwen3-Instruct-4B.Q3KL.gguf | Q3KL | 2.24 GB | | Kepler-186f-Qwen3-Instruct-4B.Q3KM.gguf | Q3KM | 2.08 GB | | Kepler-186f-Qwen3-Instruct-4B.Q3KS.gguf | Q3KS | 1.89 GB | | Kepler-186f-Qwen3-Instruct-4B.Q4KM.gguf | Q4KM | 2.5 GB | | Kepler-186f-Qwen3-Instruct-4B.Q4KS.gguf | Q4KS | 2.38 GB | | Kepler-186f-Qwen3-Instruct-4B.Q5KM.gguf | Q5KM | 2.89 GB | | Kepler-186f-Qwen3-Instruct-4B.Q5KS.gguf | Q5KS | 2.82 GB | | Kepler-186f-Qwen3-Instruct-4B.Q6K.gguf | Q6K | 3.31 GB | | Kepler-186f-Qwen3-Instruct-4B.Q80.gguf | Q80 | 4.28 GB | (sorted by size, not necessarily quality. IQ-quants are often preferable over similar sized non-IQ quants) Here is a handy graph by ikawrakow comparing some lower-quality quant types (lower is better):

> Qwen3-VL-30B-A3B-Instruct-abliterated is an abliterated (v1.0) variant of Qwen3-VL-30B-A3B-Instruct, designed for Abliterated Reasoning and Captioning. > This model leverages the Qwen3-VL-MoE (Mixture of Experts) architecture to deliver deeply descriptive, context-rich, and reasoning-oriented multimodal outputs. It handles complex, sensitive, and nuanced visual content while maintaining balanced interpretive coherence and multilingual adaptability. Abliterated / Uncensored Captioning and Reasoning Fine-tuned to bypass standard content filters while preserving factual accuracy, descriptive depth, and logical reasoning. High-Fidelity Reasoning and Visual Understanding Generates detailed captions and structured reasoning for diverse visual categories—artistic, technical, abstract, or low-context. Mixture of Experts (MoE) Efficiency Built on Qwen3-VL-MoE, dynamically routing computation through specialized experts for enhanced precision and scalability. Aspect-Ratio Robustness Performs consistently across wide, tall, square, panoramic, and irregular visual formats. Variational Detail Control Supports both concise summaries and highly detailed reasoning narratives, depending on prompt configuration. Multilingual Output Capability Defaults to English but adaptable for multilingual use through prompt engineering. Generating detailed, uncensored captions and reasoning for complex or creative visual datasets. Research in multimodal reasoning, safety evaluation, and content moderation studies. Enabling descriptive captioning and analytical reasoning for datasets excluded from mainstream models. Creative applications such as narrative generation, artistic interpretation, and visual storytelling. Advanced reasoning over diverse visual structures and aspect ratios. May produce explicit, sensitive, or offensive content depending on input and prompt. Not recommended for deployment in production systems that require strict moderation or filtering. Style, tone, and reasoning detail can vary based on prompt phrasing. May show variable performance on synthetic, abstract, or highly stylized visual inputs.

> Deepfake-vs-Real-8000 is an image classification vision-language encoder model fine-tuned from google/siglip2-base-patch16-224 for a single-label classification task. It is designed to detect whether an image is a deepfake or a real one using the SiglipForImageClassification architecture. The model categorizes images into two classes: - Class 0: "Deepfake" - Class 1: "Real one" The Deepfake-vs-Real-8000 model is designed to detect deepfake images from real ones. Potential use cases include: - Deepfake Detection: Assisting cybersecurity experts and forensic teams in detecting synthetic media. - Media Verification: Helping journalists and fact-checkers verify the authenticity of images. - AI Ethics & Research: Contributing to studies on AI-generated content detection. - Social Media Moderation: Enhancing tools to prevent misinformation and digital deception.

> QUASAR is a 4B-parameter language model fine-tuned from Qwen3-4B-Instruct-2507 through supervised learning followed by agentic reinforcement learning with tool-augmented feedback. Specially designed for generating OpenQASM 3.0 quantum circuits for tasks like QAOA and VQE, the model optimizes for both syntactic validity and semantic fidelity, using external quantum simulation for reward calculation across hierarchical criteria (syntax, distribution alignment, expectation value, and optimization progress). QUASAR is best suited for natural language to quantum circuit generation and quantum optimization algorithm design in research or integration scenarios, though users are advised to validate its outputs using external quantum simulators to address limitations in problem generalization. Training used a dataset with QASM 3.0 circuits and quantum optimization problems, employing SFT and RL (with GRPO and hierarchical reward). In evaluations, the model substantially outperforms comparable baselines—achieving leading results in syntactic correctness, distributional alignment, expectation-value matching, and high-quality circuit yield in both Pass@1 and Pass@10 metrics. | File Name | Quant Type | File Size | | - | - | - | | rlquantum4b.BF16.gguf | BF16 | 8.05 GB | | rlquantum4b.F16.gguf | F16 | 8.05 GB | | rlquantum4b.F32.gguf | F32 | 16.1 GB | | rlquantum4b.Q2K.gguf | Q2K | 1.67 GB | | rlquantum4b.Q3KL.gguf | Q3KL | 2.24 GB | | rlquantum4b.Q3KM.gguf | Q3KM | 2.08 GB | | rlquantum4b.Q3KS.gguf | Q3KS | 1.89 GB | | rlquantum4b.Q4KM.gguf | Q4KM | 2.5 GB | | rlquantum4b.Q4KS.gguf | Q4KS | 2.38 GB | | rlquantum4b.Q5KM.gguf | Q5KM | 2.89 GB | | rlquantum4b.Q5KS.gguf | Q5KS | 2.82 GB | | rlquantum4b.Q6K.gguf | Q6K | 3.31 GB | | rlquantum4b.Q80.gguf | Q80 | 4.28 GB | (sorted by size, not necessarily quality. IQ-quants are often preferable over similar sized non-IQ quants) Here is a handy graph by ikawrakow comparing some lower-quality quant types (lower is better):

> Flood-Image-Detection is a vision-language encoder model fine-tuned from `google/siglip2-base-patch16-512` for binary image classification. It is trained to detect whether an image contains a flooded scene or non-flooded environment. The model uses the `SiglipForImageClassification` architecture. > [!note] SigLIP 2: Multilingual Vision-Language Encoders with Improved Semantic Understanding, Localization, and Dense Features : https://arxiv.org/pdf/2502.14786 Disaster Monitoring – Rapid detection of flood-affected areas from imagery. Environmental Analysis – Track flooding patterns across regions using image datasets. Crisis Response – Assist emergency services in identifying critical zones. Surveillance and Safety – Monitor infrastructure or locations for flood exposure. Smart Alert Systems – Integrate with IoT or camera feeds for automated flood alerts.

> GRAM-LLaMA3.2-3B-RewardModel is a generative reward model fine-tuned from the Llama-3.2-3B-Instruct base model released by NiuTrans. It is designed to improve reward generalization for large language models (LLMs) by leveraging a novel training approach that first pre-trains on large unlabeled datasets and then fine-tunes using supervised labeled data. The training uses label smoothing and optimizes a regularized ranking loss, bridging generative and discriminative reward modeling techniques. This enables the model to be applied flexibly across a variety of tasks without the usual need for extensive fine-tuning on task-specific datasets. > GRAM-LLaMA3.2-3B-RewardModel is evaluated on the JudgeBench benchmark, which covers domains such as Chat, Code, Math, and Safety. It achieves a competitive average score of 69.9 across these categories, demonstrating strong capability for use as an open-source plug-and-play reward model that can align LLMs effectively without retraining reward models from scratch. The repository includes usage examples that let users directly apply this reward model for assessing and ranking the quality of AI-generated responses in an impartial manner. | Model File name | Size | QuantType | |---|---|---| | GRAM-LLaMA3.2-3B-RewardModel.BF16.gguf | 6.43 GB | BF16 | | GRAM-LLaMA3.2-3B-RewardModel.F16.gguf | 6.43 GB | F16 | | GRAM-LLaMA3.2-3B-RewardModel.F32.gguf | 12.9 GB | F32 | | GRAM-LLaMA3.2-3B-RewardModel.Q2K.gguf | 1.36 GB | Q2K | | GRAM-LLaMA3.2-3B-RewardModel.Q3KL.gguf | 1.82 GB | Q3KL | | GRAM-LLaMA3.2-3B-RewardModel.Q3KM.gguf | 1.69 GB | Q3KM | | GRAM-LLaMA3.2-3B-RewardModel.Q3KS.gguf | 1.54 GB | Q3KS | | GRAM-LLaMA3.2-3B-RewardModel.Q4KM.gguf | 2.02 GB | Q4KM | | GRAM-LLaMA3.2-3B-RewardModel.Q4KS.gguf | 1.93 GB | Q4KS | | GRAM-LLaMA3.2-3B-RewardModel.Q5KM.gguf | 2.32 GB | Q5KM | | GRAM-LLaMA3.2-3B-RewardModel.Q5KS.gguf | 2.27 GB | Q5KS | | GRAM-LLaMA3.2-3B-RewardModel.Q6K.gguf | 2.64 GB | Q6K | | GRAM-LLaMA3.2-3B-RewardModel.Q80.gguf | 3.42 GB | Q80 | (sorted by size, not necessarily quality. IQ-quants are often preferable over similar sized non-IQ quants) Here is a handy graph by ikawrakow comparing some lower-quality quant types (lower is better):

> Smoothie-Qwen3 models are enhancements of Qwen3 language models, applying post-processing techniques to smooth token distributions and promote balanced, multilingual output, especially across varied Unicode ranges. These models are particularly effective for applications needing reduced language bias and improved representation consistency, maintaining the strong reasoning, coding, and dialogue abilities of Qwen3 while producing more stable and diverse generations. | File Name | Quant Type | File Size | | - | - | - | | Smoothie-Qwen3-0.6B.BF16.gguf | BF16 | 1.2 GB | | Smoothie-Qwen3-0.6B.F16.gguf | F16 | 1.2 GB | | Smoothie-Qwen3-0.6B.F32.gguf | F32 | 2.39 GB | | Smoothie-Qwen3-0.6B.Q2K.gguf | Q2K | 296 MB | | Smoothie-Qwen3-0.6B.Q3KL.gguf | Q3KL | 368 MB | | Smoothie-Qwen3-0.6B.Q3KM.gguf | Q3KM | 347 MB | | Smoothie-Qwen3-0.6B.Q3KS.gguf | Q3KS | 323 MB | | Smoothie-Qwen3-0.6B.Q40.gguf | Q40 | 382 MB | | Smoothie-Qwen3-0.6B.Q41.gguf | Q41 | 409 MB | | Smoothie-Qwen3-0.6B.Q4K.gguf | Q4K | 397 MB | | Smoothie-Qwen3-0.6B.Q4KM.gguf | Q4KM | 397 MB | | Smoothie-Qwen3-0.6B.Q4KS.gguf | Q4KS | 383 MB | | Smoothie-Qwen3-0.6B.Q50.gguf | Q50 | 437 MB | | Smoothie-Qwen3-0.6B.Q51.gguf | Q51 | 464 MB | | Smoothie-Qwen3-0.6B.Q5K.gguf | Q5K | 444 MB | | Smoothie-Qwen3-0.6B.Q5KM.gguf | Q5KM | 444 MB | | Smoothie-Qwen3-0.6B.Q5KS.gguf | Q5KS | 437 MB | | Smoothie-Qwen3-0.6B.Q6K.gguf | Q6K | 495 MB | | Smoothie-Qwen3-0.6B.Q80.gguf | Q80 | 639 MB | | File Name | Quant Type | File Size | | - | - | - | | Smoothie-Qwen3-1.7B.BF16.gguf | BF16 | 3.45 GB | | Smoothie-Qwen3-1.7B.F16.gguf | F16 | 3.45 GB | | Smoothie-Qwen3-1.7B.F32.gguf | F32 | 6.89 GB | | Smoothie-Qwen3-1.7B.Q2K.gguf | Q2K | 778 MB | | Smoothie-Qwen3-1.7B.Q3KL.gguf | Q3KL | 1 GB | | Smoothie-Qwen3-1.7B.Q3KM.gguf | Q3KM | 940 MB | | Smoothie-Qwen3-1.7B.Q3KS.gguf | Q3KS | 867 MB | | Smoothie-Qwen3-1.7B.Q40.gguf | Q40 | 1.05 GB | | Smoothie-Qwen3-1.7B.Q41.gguf | Q41 | 1.14 GB | | Smoothie-Qwen3-1.7B.Q4K.gguf | Q4K | 1.11 GB | | Smoothie-Qwen3-1.7B.Q4KM.gguf | Q4KM | 1.11 GB | | Smoothie-Qwen3-1.7B.Q4KS.gguf | Q4KS | 1.06 GB | | Smoothie-Qwen3-1.7B.Q50.gguf | Q50 | 1.23 GB | | Smoothie-Qwen3-1.7B.Q51.gguf | Q51 | 1.32 GB | | Smoothie-Qwen3-1.7B.Q5K.gguf | Q5K | 1.26 GB | | Smoothie-Qwen3-1.7B.Q5KM.gguf | Q5KM | 1.26 GB | | Smoothie-Qwen3-1.7B.Q5KS.gguf | Q5KS | 1.23 GB | | Smoothie-Qwen3-1.7B.Q6K.gguf | Q6K | 1.42 GB | | Smoothie-Qwen3-1.7B.Q80.gguf | Q80 | 1.83 GB | | File Name | Quant Type | File Size | | - | - | - | | Smoothie-Qwen3-4B.BF16.gguf | BF16 | 8.05 GB | | Smoothie-Qwen3-4B.F16.gguf | F16 | 8.05 GB | | Smoothie-Qwen3-4B.F32.gguf | F32 | 16.1 GB | | Smoothie-Qwen3-4B.Q2K.gguf | Q2K | 1.67 GB | | Smoothie-Qwen3-4B.Q3KL.gguf | Q3KL | 2.24 GB | | Smoothie-Qwen3-4B.Q3KM.gguf | Q3KM | 2.08 GB | | Smoothie-Qwen3-4B.Q3KS.gguf | Q3KS | 1.89 GB | | Smoothie-Qwen3-4B.Q4KM.gguf | Q4KM | 2.5 GB | | Smoothie-Qwen3-4B.Q4KS.gguf | Q4KS | 2.38 GB | | Smoothie-Qwen3-4B.Q5KM.gguf | Q5KM | 2.89 GB | | Smoothie-Qwen3-4B.Q5KS.gguf | Q5KS | 2.82 GB | | Smoothie-Qwen3-4B.Q6K.gguf | Q6K | 3.31 GB | | Smoothie-Qwen3-4B.Q80.gguf | Q80 | 4.28 GB | (sorted by size, not necessarily quality. IQ-quants are often preferable over similar sized non-IQ quants) Here is a handy graph by ikawrakow comparing some lower-quality quant types (lower is better):

> Carinae-Qwen3-Radiation-4B is a reasoning-focused model fine-tuned on Qwen for Abliterated Reasoning and polished token probabilities, enhancing balanced multilingual generation across mathematics and general-purpose reasoning. > It specializes in event-driven logic, structured analysis, and precise probabilistic modeling—making it an ideal tool for researchers, educators, and developers working with uncertainty and structured reasoning. | File Name | Quant Type | File Size | | - | - | - | | Carinae-Qwen3-Radiation-4B.BF16.gguf | BF16 | 8.05 GB | | Carinae-Qwen3-Radiation-4B.F16.gguf | F16 | 8.05 GB | | Carinae-Qwen3-Radiation-4B.F32.gguf | F32 | 16.1 GB | | Carinae-Qwen3-Radiation-4B.Q2K.gguf | Q2K | 1.67 GB | | Carinae-Qwen3-Radiation-4B.Q3KL.gguf | Q3KL | 2.24 GB | | Carinae-Qwen3-Radiation-4B.Q3KM.gguf | Q3KM | 2.08 GB | | Carinae-Qwen3-Radiation-4B.Q3KS.gguf | Q3KS | 1.89 GB | | Carinae-Qwen3-Radiation-4B.Q4KM.gguf | Q4KM | 2.5 GB | | Carinae-Qwen3-Radiation-4B.Q4KS.gguf | Q4KS | 2.38 GB | | Carinae-Qwen3-Radiation-4B.Q5KM.gguf | Q5KM | 2.89 GB | | Carinae-Qwen3-Radiation-4B.Q5KS.gguf | Q5KS | 2.82 GB | | Carinae-Qwen3-Radiation-4B.Q6K.gguf | Q6K | 3.31 GB | | Carinae-Qwen3-Radiation-4B.Q80.gguf | Q80 | 4.28 GB | (sorted by size, not necessarily quality. IQ-quants are often preferable over similar sized non-IQ quants) Here is a handy graph by ikawrakow comparing some lower-quality quant types (lower is better):

> Face-Mask-Detection is a binary image classification model based on `google/siglip2-base-patch16-224`, trained to detect whether a person is wearing a face mask or not. This model can be used in public health monitoring, access control systems, and workplace compliance enforcement. The model distinguishes between the following face mask statuses: COVID-19 Compliance Monitoring Security and Access Control Automated Surveillance Systems Health Safety Enforcement in Public Spaces

| Parameter | Value | Parameter | Value | |---------------------------|--------|---------------------------|--------| | LR Scheduler | constant | Noise Offset | 0.03 | | Optimizer | AdamW | Multires Noise Discount | 0.1 | | Network Dim | 64 | Multires Noise Iterations | 10 | | Network Alpha | 32 | Repeat & Steps | 25 & 3000 | | Epoch | 20 | Save Every N Epochs | 1 | | Source | Link | |--------------|-------------------------------------| | Playground | playground.com | | ArtStation | artstation.com | | 4K Wallpapers| 4kwallpapers.com | | Dimensions | Aspect Ratio | Recommendation | |-----------------|------------------|---------------------------| | 1664 x 928 | 16:9 (approx.) | Best | | 1024 x 1024 | 1:1 | Default | You should use `Qwen Anime` to trigger the image generation.

> Gliese-QueryTool-0.6B is a function-calling and query-oriented reasoning model fine-tuned from Qwen3-0.6B using Salesforce/xlam-function-calling-60k, designed for tool orchestration, structured query resolution, and operation chaining across diverse tasks. It excels in dynamic function execution, structured reasoning pipelines, and multi-tool decision workflows, making it a powerful lightweight solution for developers, tooling platforms, and automation systems. Model Files | File Name | Quant Type | File Size | | - | - | - | | Gliese-QueryTool-0.6B.BF16.gguf | BF16 | 1.2 GB | | Gliese-QueryTool-0.6B.F16.gguf | F16 | 1.2 GB | | Gliese-QueryTool-0.6B.F32.gguf | F32 | 2.39 GB | | Gliese-QueryTool-0.6B.Q2K.gguf | Q2K | 296 MB | | Gliese-QueryTool-0.6B.Q3KL.gguf | Q3KL | 368 MB | | Gliese-QueryTool-0.6B.Q3KM.gguf | Q3KM | 347 MB | | Gliese-QueryTool-0.6B.Q3KS.gguf | Q3KS | 323 MB | | Gliese-QueryTool-0.6B.Q40.gguf | Q40 | 382 MB | | Gliese-QueryTool-0.6B.Q41.gguf | Q41 | 409 MB | | Gliese-QueryTool-0.6B.Q4K.gguf | Q4K | 397 MB | | Gliese-QueryTool-0.6B.Q4KM.gguf | Q4KM | 397 MB | | Gliese-QueryTool-0.6B.Q4KS.gguf | Q4KS | 383 MB | | Gliese-QueryTool-0.6B.Q50.gguf | Q50 | 437 MB | | Gliese-QueryTool-0.6B.Q51.gguf | Q51 | 464 MB | | Gliese-QueryTool-0.6B.Q5K.gguf | Q5K | 444 MB | | Gliese-QueryTool-0.6B.Q5KM.gguf | Q5KM | 444 MB | | Gliese-QueryTool-0.6B.Q5KS.gguf | Q5KS | 437 MB | | Gliese-QueryTool-0.6B.Q6K.gguf | Q6K | 495 MB | | Gliese-QueryTool-0.6B.Q80.gguf | Q80 | 639 MB | (sorted by size, not necessarily quality. IQ-quants are often preferable over similar sized non-IQ quants) Here is a handy graph by ikawrakow comparing some lower-quality quant types (lower is better):

> Qwen3-VL-2B-Instruct-abliterated is an abliterated (v1.0) variant of Qwen3-VL-2B-Instruct, designed for Abliterated Reasoning and Captioning. > This model is optimized to generate detailed, descriptive captions and reasoning outputs across a wide range of visual and multimodal contexts—including complex, sensitive, or nuanced content—while supporting diverse aspect ratios and resolutions. Abliterated / Uncensored Captioning – Fine-tuned to bypass conventional content filters while preserving factual, descriptive, and reasoning-rich outputs. High-Fidelity Descriptions – Generates comprehensive captions and reasoning for general, artistic, technical, abstract, or low-context images. Robust Across Aspect Ratios – Performs consistently across wide, tall, square, and irregular image dimensions. Variational Detail Control – Capable of producing outputs ranging from concise summaries to fine-grained, intricate descriptions and reasoning. Foundation on Qwen3-VL-2B Architecture – Built upon Qwen3-VL-2B-Instruct’s strong multimodal reasoning and instruction-following capabilities. Multilingual Output Capability – Primarily optimized for English, with adaptability for multilingual prompts through prompt engineering. Generating detailed, uncensored captions and reasoning for general-purpose or artistic datasets. Research in content moderation, red-teaming, and generative safety evaluation. Enabling descriptive captioning and reasoning for visual datasets typically excluded from mainstream models. Creative applications such as storytelling, art generation, or multimodal reasoning tasks. Captioning and reasoning for non-standard aspect ratios and stylized visual content. May produce explicit, sensitive, or offensive descriptions depending on the image content and prompts. Not recommended for production systems requiring strict content moderation. Output style, tone, and reasoning may vary based on input phrasing. Accuracy can fluctuate for unfamiliar, synthetic, or highly abstract visual content.

> GRAM-Qwen3-4B-RewardModel is a generative reward model developed to address reward generalization for Large Language Models (LLMs), released by NiuTrans. Unlike traditional models that depend heavily on task-specific labeled data, this model leverages both labeled and unlabeled data—a novel approach that allows it to generalize better across various tasks. It introduces a generative reward model framework that pre-trains on large amounts of unlabeled data and is subsequently fine-tuned with supervised data. The methodology also employs label smoothing and a regularized ranking loss to further boost performance, effectively bridging the gap between generative and discriminative reward modeling techniques. > This model is built on the Qwen3-4B base and can be directly used or adapted for aligning LLMs without the need to train a reward model from scratch on extensive datasets. In evaluations on the JudgeBench benchmark—covering Chat, Code, Math, and Safety tasks—GRAM-Qwen3-4B-RewardModel achieves a competitive average score of 65.9, making it suitable for use as an open-source, plug-and-play reward model for a variety of LLM alignment scenarios. The repository provides usage instructions and demonstration code to facilitate immediate adoption for research and development purposes | Model File name | Size | QuantType | |---|---|---| | GRAM-Qwen3-4B-RewardModel.BF16.gguf | 8.05 GB | BF16 | | GRAM-Qwen3-4B-RewardModel.F16.gguf | 8.05 GB | F16 | | GRAM-Qwen3-4B-RewardModel.F32.gguf | 16.1 GB | F32 | | GRAM-Qwen3-4B-RewardModel.Q2K.gguf | 1.67 GB | Q2K | | GRAM-Qwen3-4B-RewardModel.Q3KL.gguf | 2.24 GB | Q3KL | | GRAM-Qwen3-4B-RewardModel.Q3KM.gguf | 2.08 GB | Q3KM | | GRAM-Qwen3-4B-RewardModel.Q3KS.gguf | 1.89 GB | Q3KS | | GRAM-Qwen3-4B-RewardModel.Q4KM.gguf | 2.5 GB | Q4KM | | GRAM-Qwen3-4B-RewardModel.Q4KS.gguf | 2.38 GB | Q4KS | | GRAM-Qwen3-4B-RewardModel.Q5KM.gguf | 2.89 GB | Q5KM | | GRAM-Qwen3-4B-RewardModel.Q5KS.gguf | 2.82 GB | Q5KS | | GRAM-Qwen3-4B-RewardModel.Q6K.gguf | 3.31 GB | Q6K | | GRAM-Qwen3-4B-RewardModel.Q80.gguf | 4.28 GB | Q80 | (sorted by size, not necessarily quality. IQ-quants are often preferable over similar sized non-IQ quants) Here is a handy graph by ikawrakow comparing some lower-quality quant types (lower is better):

> Mirage-Photo-Classifier is an image classification vision-language encoder model fine-tuned from google/siglip2-base-patch16-224 for a binary image authenticity classification task. It is designed to determine whether an image is real or AI-generated (fake) using the SiglipForImageClassification architecture. The Mirage-Photo-Classifier model is designed to detect whether an image is genuine (photograph) or synthetically generated. Use cases include: - AI Image Detection: Identifying AI-generated images in social media, news, or datasets. - Digital Forensics: Helping professionals detect image authenticity in investigations. - Platform Moderation: Assisting content platforms in labeling generated content. - Dataset Validation: Cleaning and verifying training data for other AI models.

> GRAM-Qwen3-1.7B-RewardModel is a generative reward model developed by NiuTrans that follows a two-step training approach: it first pre-trains on a large amount of unlabeled data and then fine-tunes with supervised labeled data. This methodology, which incorporates label smoothing and a regularized ranking loss, enables effective reward generalization for large language models (LLMs). The model is built on the Qwen3-1.7B base, a compact language model with 1.7 billion parameters, 28 layers, and attention heads designed to handle long-context inputs (up to 32,768 tokens) and support both detailed reasoning and fast responses. GRAM-Qwen3-1.7B-RewardModel is intended for flexible application across diverse tasks, providing an open-source, plug-and-play reward model for aligning LLM outputs without requiring extensive task-specific retraining. It excels in evaluating and ranking the quality of AI-generated responses, operating effectively as a judge model in AI alignment scenarios. | Model File name | Size | QuantType | |---|---|---| | GRAM-Qwen3-1.7B-RewardModel.BF16.gguf | 3.45 GB | BF16 | | GRAM-Qwen3-1.7B-RewardModel.F16.gguf | 3.45 GB | F16 | | GRAM-Qwen3-1.7B-RewardModel.F32.gguf | 6.89 GB | F32 | | GRAM-Qwen3-1.7B-RewardModel.Q2K.gguf | 778 MB | Q2K | | GRAM-Qwen3-1.7B-RewardModel.Q3KL.gguf | 1 GB | Q3KL | | GRAM-Qwen3-1.7B-RewardModel.Q3KM.gguf | 940 MB | Q3KM | | GRAM-Qwen3-1.7B-RewardModel.Q3KS.gguf | 867 MB | Q3KS | | GRAM-Qwen3-1.7B-RewardModel.Q4KM.gguf | 1.11 GB | Q4KM | | GRAM-Qwen3-1.7B-RewardModel.Q4KS.gguf | 1.06 GB | Q4KS | | GRAM-Qwen3-1.7B-RewardModel.Q5KM.gguf | 1.26 GB | Q5KM | | GRAM-Qwen3-1.7B-RewardModel.Q5KS.gguf | 1.23 GB | Q5KS | | GRAM-Qwen3-1.7B-RewardModel.Q6K.gguf | 1.42 GB | Q6K | | GRAM-Qwen3-1.7B-RewardModel.Q80.gguf | 1.83 GB | Q80 | (sorted by size, not necessarily quality. IQ-quants are often preferable over similar sized non-IQ quants) Here is a handy graph by ikawrakow comparing some lower-quality quant types (lower is better):

> Qwen3-VL-2B-Thinking-abliterated is an abliterated (v1.0) variant of Qwen3-VL-2B-Thinking, designed for Abliterated Reasoning and Captioning. > This model is optimized to generate detailed, descriptive captions and reasoning outputs across a wide range of visual and multimodal contexts—including complex, sensitive, or nuanced content—while supporting diverse aspect ratios and resolutions. Abliterated / Uncensored Captioning – Fine-tuned to bypass conventional content filters while preserving factual, descriptive, and reasoning-rich outputs. High-Fidelity Descriptions – Generates comprehensive captions and reasoning for general, artistic, technical, abstract, or low-context images. Robust Across Aspect Ratios – Performs consistently across wide, tall, square, and irregular image dimensions. Variational Detail Control – Capable of producing outputs ranging from concise summaries to fine-grained, intricate descriptions and reasoning. Foundation on Qwen3-VL-2B-Thinking Architecture – Built upon Qwen3-VL-2B-Thinking’s strong multimodal reasoning and instruction-following capabilities. Multilingual Output Capability – Primarily optimized for English, with adaptability for multilingual prompts through prompt engineering. Generating detailed, uncensored captions and reasoning for general-purpose or artistic datasets. Research in content moderation, red-teaming, and generative safety evaluation. Enabling descriptive captioning and reasoning for visual datasets typically excluded from mainstream models. Creative applications such as storytelling, art generation, or multimodal reasoning tasks. Captioning and reasoning for non-standard aspect ratios and stylized visual content. May produce explicit, sensitive, or offensive descriptions depending on the image content and prompts. Not recommended for production systems requiring strict content moderation. Output style, tone, and reasoning may vary based on input phrasing. Accuracy can fluctuate for unfamiliar, synthetic, or highly abstract visual content.

> Qwen3-VL-8B-Thinking-abliterated is an abliterated (v1.0) variant of Qwen3-VL-8B-Thinking, designed for Abliterated Reasoning and Captioning. > This model produces detailed captions and reasoning outputs across a wide range of visual and multimodal contexts, including complex, sensitive, or nuanced content. It supports diverse aspect ratios, resolutions, and prompt conditions while maintaining reasoning integrity and descriptive precision. Abliterated / Uncensored Captioning Fine-tuned to bypass conventional content filters while preserving factual, descriptive, and reasoning-rich outputs. High-Fidelity Reasoning and Descriptions Generates comprehensive captions and reasoning for general, artistic, technical, abstract, and low-context images. Robust Across Aspect Ratios Performs consistently on wide, tall, square, panoramic, and irregular image dimensions. Variational Detail Control Produces outputs ranging from concise summaries to fine-grained, high-context reasoning and descriptions. Foundation on Qwen3-VL-8B-Thinking Architecture Built upon the Qwen3-VL-8B-Thinking model’s advanced multimodal reasoning and instruction-following capabilities. Multilingual Output Capability Defaults to English but can be adapted to other languages via prompt engineering. Generating detailed, uncensored captions and reasoning for general-purpose, artistic, or research-oriented datasets. Research in content moderation, red-teaming, and generative safety analysis. Enabling descriptive captioning and reasoning for datasets typically excluded from mainstream models. Creative applications such as visual storytelling, art description, and multimodal reasoning exploration. Captioning and reasoning for images with non-standard or stylized visual structures. May generate explicit, sensitive, or offensive content depending on prompts and image input. Not suitable for production systems that require strict content moderation. Output style, tone, and reasoning depth may vary based on input phrasing. Accuracy may fluctuate for abstract, synthetic, or highly stylized visuals.

> GiD-Land-Cover-Classification is a multi-class image classification model based on `google/siglip2-base-patch16-224`, trained to detect land cover types in geographical or environmental imagery. This model can be used for urban planning, agriculture monitoring, and environmental analysis. The model distinguishes between the following land cover types: Urban Development Planning Agricultural Monitoring Land Use and Land Cover (LULC) Mapping Disaster Management and Flood Risk Analysis

> Dots.OCR-Latest-BF16 is an optimized and updated vision-language OCR model variant of the original Dots.OCR. This open-source model is designed to extract text from images and scanned documents, including handwritten and printed content. It can output results as plain text or Markdown, preserving document layout elements such as headings, tables, and lists. This model uses a powerful multimodal backbone (3B VLM) to enhance reading comprehension and layout understanding, handling cursive handwriting and complex document structures effectively. The BF16 variant has been tested and updated to work smoothly with the latest `transformers` version without compatibility issues, ensuring optimized performance. | Resource Type | Description | Link | |----------------|--------------|------| | Original Model Card | Official release of Dots.OCR by rednote-hilab | rednote-hilab/dots.ocr | | Test Model (StrangerZone HF) | Community test deployment (experimental) | strangervisionhf/dots.ocr-base-fix | | Standard Model Card | Optimized version supporting Transformers v4.57.1 (BF16 precision) | prithivMLmods/Dots.OCR-Latest-BF16 | | Demo Space | Interactive demo hosted on Hugging Face Spaces | Multimodal-OCR3 Demo |

> Light-IF-4B is a 4-billion-parameter instruction-following language model derived from Qwen3-4B-Base, designed to overcome "lazy reasoning" in complex tasks by incorporating previewing and self-checking during inference; it is fine-tuned using entropy-preserving supervised learning (Entropy-SFT) and token-wise entropy-adaptive reinforcement learning (TEA-RL) on a carefully filtered dataset, producing strong results across instruction-following and reasoning benchmarks (such as SuperClue, IFEval, and IFBench), where it matches or outperforms even larger or closed-source models, and supports advanced features such as extended context (32k-131k tokens with YaRN), efficient deployment (via Hugging Face Transformers, sglang, or vllm), and open integration for research in robust generalizable reasoning, with further details, evaluation code, and licensing on its official Hugging Face repository and paper. | File Name | Size | Quant Type | |-----------|------|------------| | Qwen3-4B-MegaScience.BF16.gguf | 8.05 GB | BF16 | | Qwen3-4B-MegaScience.F16.gguf | 8.05 GB | F16 | | Qwen3-4B-MegaScience.F32.gguf | 16.1 GB | F32 | | Qwen3-4B-MegaScience.Q3KL.gguf | 2.24 GB | Q3KL | | Qwen3-4B-MegaScience.Q3KS.gguf | 1.89 GB | Q3KS | | Qwen3-4B-MegaScience.Q4KM.gguf | 2.5 GB | Q4KM | | Qwen3-4B-MegaScience.Q4KS.gguf | 2.38 GB | Q4KS | | Qwen3-4B-MegaScience.Q5KM.gguf | 2.89 GB | Q5KM | | Qwen3-4B-MegaScience.Q5KS.gguf | 2.82 GB | Q5KS | | Qwen3-4B-MegaScience.Q6K.gguf | 3.31 GB | Q6K | | Qwen3-4B-MegaScience.Q80.gguf | 4.28 GB | Q80 | (sorted by size, not necessarily quality. IQ-quants are often preferable over similar sized non-IQ quants) Here is a handy graph by ikawrakow comparing some lower-quality quant types (lower is better):

> The Qwen3-VL-8B-Abliterated-Caption-it model is a fine-tuned version of Qwen3-VL-8B-Instruct, tailored for Abliterated Captioning / Uncensored Image Captioning. This variant is designed to generate highly detailed and descriptive captions across a broad range of visual categories, including images with complex, sensitive, or nuanced content—across varying aspect ratios and resolutions. Abliterated / Uncensored Captioning: Fine-tuned to bypass common content filters while preserving factual and descriptive richness across diverse visual categories. High-Fidelity Descriptions: Generates comprehensive captions for general, artistic, technical, abstract, and low-context images. Robust Across Aspect Ratios: Capable of accurately captioning images with wide, tall, square, and irregular dimensions. Variational Detail Control: Produces outputs with both high-level summaries and fine-grained descriptions as needed. Foundation on Qwen3-VL Architecture: Leverages the strengths of the Qwen3-VL-8B multimodal model for visual reasoning, comprehension, and instruction-following. Multilingual Output Capability: Supports multilingual descriptions (English as default), adaptable via prompt engineering. > [!note] > Instruction Query: Provide a detailed caption for the image Generating detailed and unfiltered image captions for general-purpose or artistic datasets. Content moderation research, red-teaming, and generative safety evaluations. Enabling descriptive captioning for visual datasets typically excluded from mainstream models. Creative applications (e.g., storytelling, art generation) that benefit from rich descriptive captions. Captioning for non-standard aspect ratios and stylized visual content. May produce explicit, sensitive, or offensive descriptions depending on image content and prompts. Not suitable for deployment in production systems requiring content filtering or moderation. Can exhibit variability in caption tone or style depending on input prompt phrasing. Accuracy for unfamiliar or synthetic visual styles may vary.

> II-Search-CIR-4B is a 4-billion-parameter language model built on Qwen3-4B and enhanced with Code-Integrated Reasoning (CIR), enabling it not only to call external tools (such as web search and web visit) through code blocks during inference, but also to programmatically process, filter, and reason over results within those code blocks; optimized through supervised fine-tuning and reinforcement learning on challenging reasoning datasets, the model achieves state-of-the-art or leading results on major factual QA and information-seeking benchmarks (like OpenAI/SimpleQA, Google/Frames, and Seal0), and it can be efficiently deployed using vLLM or SGLang with up to 128k-token contexts (with YaRN RoPE scaling), supporting advanced research, educational, and web-integrated applications, with datasets, code samples, and evaluation results provided in the official Hugging Face repository. | File Name | Size | Quant Type | |-----------|------|------------| | II-Search-4B-GGUF.BF16.gguf | 8.05 GB | BF16 | | II-Search-4B-GGUF.F16.gguf | 8.05 GB | F16 | | II-Search-4B-GGUF.F32.gguf | 16.1 GB | F32 | | II-Search-4B-GGUF.Q2K.gguf | 1.67 GB | Q2K | | II-Search-4B-GGUF.Q3KL.gguf | 2.24 GB | Q3KL | | II-Search-4B-GGUF.Q3KM.gguf | 2.08 GB | Q3KM | | II-Search-4B-GGUF.Q3KS.gguf | 1.89 GB | Q3KS | | II-Search-4B-GGUF.Q4KM.gguf | 2.5 GB | Q4KM | | II-Search-4B-GGUF.Q4KS.gguf | 2.38 GB | Q4KS | | II-Search-4B-GGUF.Q5KM.gguf | 2.89 GB | Q5KM | | II-Search-4B-GGUF.Q5KS.gguf | 2.82 GB | Q5KS | | II-Search-4B-GGUF.Q6K.gguf | 3.31 GB | Q6K | | II-Search-4B-GGUF.Q80.gguf | 4.28 GB | Q80 | (sorted by size, not necessarily quality. IQ-quants are often preferable over similar sized non-IQ quants) Here is a handy graph by ikawrakow comparing some lower-quality quant types (lower is better):

> The Qwen2-VL-2B-Abliterated-Caption-it model is a fine-tuned version of Qwen2-VL-2B-Instruct, tailored for Abliterated Captioning / Uncensored Image Captioning. This variant is designed to generate highly detailed and descriptive captions across a broad range of visual categories, including images with complex, sensitive, or nuanced content—across varying aspect ratios and resolutions. Abliterated / Uncensored Captioning: Fine-tuned to bypass common content filters while preserving factual and descriptive richness across diverse visual categories. High-Fidelity Descriptions: Generates comprehensive captions for general, artistic, technical, abstract, and low-context images. Robust Across Aspect Ratios: Capable of accurately captioning images with wide, tall, square, and irregular dimensions. Variational Detail Control: Produces outputs with both high-level summaries and fine-grained descriptions as needed. Foundation on Qwen2-VL Architecture: Leverages the strengths of the Qwen2-VL-2B multimodal model for visual reasoning, comprehension, and instruction-following. Multilingual Output Capability: Can support multilingual descriptions (English as default), adaptable via prompt engineering. This model was fine-tuned using the following datasets: prithivMLmods/blip3o-caption-mini-arrow prithivMLmods/Caption3o-Opt-v2 prithivMLmods/Caption3o-LongCap-v4 Private/unlisted datasets curated for uncensored and domain-specific image captioning tasks. The training objective focused on enhancing performance in unconstrained, descriptive image captioning—especially for edge cases commonly filtered out in standard captioning benchmarks. > \[!note] > General Query: Caption the image precisely. | Demo | |------| | [](https://huggingface.co/prithivMLmods/Qwen2-VL-2B-Abliterated-Caption-it/blob/main/Qwen2-VL-2B-Abliterated-Caption-it/Qwen2VL2BAbliteratedCaptionit.ipynb) | Generating detailed and unfiltered image captions for general-purpose or artistic datasets. Content moderation research, red-teaming, and generative safety evaluations. Enabling descriptive captioning for visual datasets typically excluded from mainstream models. Use in creative applications (e.g., storytelling, art generation) that benefit from rich descriptive captions. Captioning for non-standard aspect ratios and stylized visual content. May produce explicit, sensitive, or offensive descriptions depending on image content and prompts. Not suitable for deployment in production systems requiring content filtering or moderation. Can exhibit variability in caption tone or style depending on input prompt phrasing. Accuracy for unfamiliar or synthetic visual styles may vary.

> WikiArt-Style is a vision model fine-tuned from google/siglip2-base-patch16-224 using the SiglipForImageClassification architecture. It classifies art images into one of 137 painting style categories. {'0': 'Abstract Art', '1': 'Abstract Expressionism', '2': 'Academicism', '3': 'Action painting', '4': 'American Realism', '5': 'Analytical Cubism', '6': 'Analytical\xa0Realism', '7': 'Art Brut', '8': 'Art Deco', '9': 'Art Informel', '10': 'Art Nouveau (Modern)', '11': 'Automatic Painting', '12': 'Baroque', '13': 'Biedermeier', '14': 'Byzantine', '15': 'Cartographic Art', '16': 'Classicism', '17': 'Cloisonnism', '18': 'Color Field Painting', '19': 'Conceptual Art', '20': 'Concretism', '21': 'Constructivism', '22': 'Contemporary Realism', '23': 'Costumbrismo', '24': 'Cubism', '25': 'Cubo-Expressionism', '26': 'Cubo-Futurism', '27': 'Dada', '28': 'Divisionism', '29': 'Early Renaissance', '30': 'Environmental (Land) Art', '31': 'Existential Art', '32': 'Expressionism', '33': 'Fantastic Realism', '34': 'Fauvism', '35': 'Feminist Art', '36': 'Figurative Expressionism', '37': 'Futurism', '38': 'Gongbi', '39': 'Gothic', '40': 'Hard Edge Painting', '41': 'High Renaissance', '42': 'Hyper-Realism', '43': 'Ilkhanid', '44': 'Impressionism', '45': 'Indian Space painting', '46': 'Ink and wash painting', '47': 'International Gothic', '48': 'Intimism', '49': 'Japonism', '50': 'Joseon Dynasty', '51': 'Kinetic Art', '52': 'Kitsch', '53': 'Lettrism', '54': 'Light and Space', '55': 'Luminism', '56': 'Lyrical Abstraction', '57': 'Magic Realism', '58': 'Mail Art', '59': 'Mannerism (Late Renaissance)', '60': 'Mechanistic Cubism', '61': 'Metaphysical art', '62': 'Minimalism', '63': 'Miserablism', '64': 'Modernismo', '65': 'Mosan art', '66': 'Muralism', '67': 'Nanga (Bunjinga)', '68': 'Nas-Taliq', '69': 'Native Art', '70': 'Naturalism', '71': 'Naïve Art (Primitivism)', '72': 'Neo-Byzantine', '73': 'Neo-Concretism', '74': 'Neo-Dada', '75': 'Neo-Expressionism', '76': 'Neo-Figurative Art', '77': 'Neo-Rococo', '78': 'Neo-Romanticism', '79': 'Neo-baroque', '80': 'Neoclassicism', '81': 'Neoplasticism', '82': 'New Casualism', '83': 'New European Painting', '84': 'New Realism', '85': 'Nihonga', '86': 'None', '87': 'Northern Renaissance', '88': 'Nouveau Réalisme', '89': 'Op Art', '90': 'Orientalism', '91': 'Orphism', '92': 'Ottoman Period', '93': 'Outsider art', '94': 'Perceptism ', '95': 'Photorealism', '96': 'Pointillism', '97': 'Pop Art', '98': 'Post-Impressionism', '99': 'Post-Minimalism', '100': 'Post-Painterly Abstraction', '101': 'Poster Art Realism', '102': 'Precisionism', '103': 'Primitivism', '104': 'Proto Renaissance', '105': 'Purism', '106': 'Rayonism', '107': 'Realism', '108': 'Regionalism', '109': 'Renaissance', '110': 'Rococo', '111': 'Romanesque', '112': 'Romanticism', '113': 'Safavid Period', '114': 'Shin-hanga', '115': 'Social Realism', '116': 'Socialist Realism', '117': 'Spatialism', '118': 'Spectralism', '119': 'Street art', '120': 'Suprematism', '121': 'Surrealism', '122': 'Symbolism', '123': 'Synchromism', '124': 'Synthetic Cubism', '125': 'Synthetism', '126': 'Sōsaku hanga', '127': 'Tachisme', '128': 'Tenebrism', '129': 'Timurid Period', '130': 'Tonalism', '131': 'Transautomatism', '132': 'Tubism', '133': 'Ukiyo-e', '134': 'Verism', '135': 'Yamato-e', '136': 'Zen'} The model predicts one of the following painting style categories: 1. Style Classification in Machine Learning Models - Used as labels for training and evaluating models that classify artworks based on their artistic styles. - Ideal for deep learning applications involving convolutional neural networks (CNNs) or transformer-based vision models. 2. Style Transfer Applications - Acts as a style reference for neural style transfer algorithms (e.g., applying "Baroque" or "Cubism" to photos). - Can guide users to select a target style from a curated list. 3. Dataset Annotation - Used to annotate images in large datasets of paintings with consistent style names. - Ensures compatibility with datasets like WikiArt, Kaggle’s Painter by Numbers, or custom curation. 4. Educational and Exploratory Interfaces - Powers interfaces or apps for exploring art history, with filterable and searchable styles. - Great for building art recommender systems or virtual museums. 5. Generative Art Prompting - Assists in text-to-image prompting for generative models (e.g., Stable Diffusion, DALL·E) to specify desired styles. - Example: "Generate a portrait in the style of Neo-Expressionism." 6. Metadata Categorization in Art Databases - Useful for tagging and organizing artworks by style in digital archives or NFT marketplaces.

> Human-Action-Recognition is an image classification vision-language encoder model fine-tuned from google/siglip2-base-patch16-224 for multi-class human action recognition. It uses the SiglipForImageClassification architecture to predict human activities from still images. The model categorizes images into 15 action classes: - 0: calling - 1: clapping - 2: cycling - 3: dancing - 4: drinking - 5: eating - 6: fighting - 7: hugging - 8: laughing - 9: listeningtomusic - 10: running - 11: sitting - 12: sleeping - 13: texting - 14: usinglaptop The Human-Action-Recognition model is designed to detect and classify human actions from images. Example applications: - Surveillance & Monitoring: Recognizing suspicious or specific activities in public spaces. - Sports Analytics: Identifying player activities or movements. - Social Media Insights: Understanding trends in user-posted visuals. - Healthcare: Monitoring elderly or patients for activity patterns. - Robotics & Automation: Enabling context-aware AI systems with visual understanding.

> Canum-med-Qwen3-Reasoning is an experimental medical reasoning and advisory model fine-tuned on Qwen/Qwen3-1.7B using the MTEB/rawmedrxiv dataset. It is designed to support clinical reasoning, biomedical understanding, and structured advisory outputs, making it a useful tool for researchers, educators, and medical professionals in experimental workflows. | File Name | Quant Type | File Size | | - | - | - | | Canum-med-Qwen3-Reasoning.BF16.gguf | BF16 | 3.45 GB | | Canum-med-Qwen3-Reasoning.F16.gguf | F16 | 3.45 GB | | Canum-med-Qwen3-Reasoning.F32.gguf | F32 | 6.89 GB | | Canum-med-Qwen3-Reasoning.Q2K.gguf | Q2K | 778 MB | | Canum-med-Qwen3-Reasoning.Q3KL.gguf | Q3KL | 1 GB | | Canum-med-Qwen3-Reasoning.Q3KM.gguf | Q3KM | 940 MB | | Canum-med-Qwen3-Reasoning.Q3KS.gguf | Q3KS | 867 MB | | Canum-med-Qwen3-Reasoning.Q40.gguf | Q40 | 1.05 GB | | Canum-med-Qwen3-Reasoning.Q41.gguf | Q41 | 1.14 GB | | Canum-med-Qwen3-Reasoning.Q4K.gguf | Q4K | 1.11 GB | | Canum-med-Qwen3-Reasoning.Q4KM.gguf | Q4KM | 1.11 GB | | Canum-med-Qwen3-Reasoning.Q4KS.gguf | Q4KS | 1.06 GB | | Canum-med-Qwen3-Reasoning.Q50.gguf | Q50 | 1.23 GB | | Canum-med-Qwen3-Reasoning.Q51.gguf | Q51 | 1.32 GB | | Canum-med-Qwen3-Reasoning.Q5K.gguf | Q5K | 1.26 GB | | Canum-med-Qwen3-Reasoning.Q5KM.gguf | Q5KM | 1.26 GB | | Canum-med-Qwen3-Reasoning.Q5KS.gguf | Q5KS | 1.23 GB | | Canum-med-Qwen3-Reasoning.Q6K.gguf | Q6K | 1.42 GB | | Canum-med-Qwen3-Reasoning.Q80.gguf | Q80 | 1.83 GB | (sorted by size, not necessarily quality. IQ-quants are often preferable over similar sized non-IQ quants) Here is a handy graph by ikawrakow comparing some lower-quality quant types (lower is better):

> Qwen3-VL-32B-Thinking-abliterated is an abliterated (v1.0) variant of Qwen3-VL-32B-Thinking, designed for Abliterated Reasoning and Captioning. > This model is optimized to generate detailed, descriptive captions and reasoning outputs across a wide range of visual and multimodal contexts—including complex, sensitive, or nuanced content—while supporting diverse aspect ratios and resolutions. Abliterated / Uncensored Captioning – Fine-tuned to bypass conventional content filters while preserving factual, descriptive, and reasoning-rich outputs. High-Fidelity Descriptions – Generates comprehensive captions and reasoning for general, artistic, technical, abstract, or low-context images. Robust Across Aspect Ratios – Performs consistently across wide, tall, square, and irregular image dimensions. Variational Detail Control – Capable of producing outputs ranging from concise summaries to fine-grained, intricate descriptions and reasoning. Foundation on Qwen3-VL-32B Architecture – Built upon Qwen3-VL-32B-Thinking’s advanced multimodal reasoning and instruction-following capabilities. Multilingual Output Capability – Primarily optimized for English, with adaptability for multilingual prompts through prompt engineering. Generating detailed, uncensored captions and reasoning for general-purpose or artistic datasets. Research in content moderation, red-teaming, and generative safety evaluation. Enabling descriptive captioning and reasoning for visual datasets typically excluded from mainstream models. Creative applications such as storytelling, art generation, or multimodal reasoning tasks. Captioning and reasoning for non-standard aspect ratios and stylized visual content. May produce explicit, sensitive, or offensive descriptions depending on the image content and prompts. Not recommended for production systems requiring strict content moderation. Output style, tone, and reasoning may vary based on input phrasing. Accuracy can fluctuate for unfamiliar, synthetic, or highly abstract visual content.

This is an ONNX version of prithivMLmods/Common-Voice-Gender-Detection. It was automatically converted and uploaded using this space. > Common-Voice-Gender-Detection is a fine-tuned version of `facebook/wav2vec2-base-960h` for binary audio classification, specifically trained to detect speaker gender as female or male. This model leverages the `Wav2Vec2ForSequenceClassification` architecture for efficient and accurate voice-based gender classification. > [!note] Wav2Vec2: Self-Supervised Learning for Speech Recognition : https://arxiv.org/pdf/2006.11477 Speech Analytics – Assist in analyzing speaker demographics in call centers or customer service recordings. Conversational AI Personalization – Adjust tone or dialogue based on gender detection for more personalized voice assistants. Voice Dataset Curation – Automatically tag or filter voice datasets by speaker gender for better dataset management. Research Applications – Enable linguistic and acoustic research involving gender-specific speech patterns. Multimedia Content Tagging – Automate metadata generation for gender identification in podcasts, interviews, or video content.

Galactic-Qwen-14B-Exp2 is based on the Qwen 2.5 14B modality architecture, designed to enhance the reasoning capabilities of 14B-parameter models. This model is optimized for general-purpose reasoning and answering, excelling in contextual understanding, logical deduction, and multi-step problem-solving. It has been fine-tuned using a long chain-of-thought reasoning model and specialized datasets to improve comprehension, structured responses, and conversational intelligence. Key Improvements 1. Enhanced General Knowledge: The model provides broad knowledge across various domains, improving capabilities in answering questions accurately and generating coherent responses. 2. Improved Instruction Following: Significant advancements in understanding and following complex instructions, generating structured responses, and maintaining coherence over extended interactions. 3. Versatile Adaptability: More resilient to diverse prompts, enhancing its ability to handle a wide range of topics and conversation styles, including open-ended and structured inquiries. 4. Long-Context Support: Supports up to 128K tokens for input context and can generate up to 8K tokens in a single output, making it ideal for detailed responses. 5. Multilingual Proficiency: Supports over 29 languages, including English, Chinese, French, Spanish, Portuguese, German, Italian, Russian, Japanese, Korean, Vietnamese, Thai, Arabic, and more. Here is a code snippet with `applychattemplate` to show you how to load the tokenizer and model and generate content: Intended Use 1. General-Purpose Reasoning: Designed for broad applicability, assisting with logical reasoning, answering diverse questions, and solving general knowledge problems. 2. Educational and Informational Assistance: Suitable for providing explanations, summaries, and research-based responses for students, educators, and general users. 3. Conversational AI and Chatbots: Ideal for building intelligent conversational agents that require contextual understanding and dynamic response generation. 4. Multilingual Applications: Supports global communication, translations, and multilingual content generation. 5. Structured Data Processing: Capable of analyzing and generating structured outputs, such as tables and JSON, useful for data science and automation. 6. Long-Form Content Generation: Can generate extended responses, including articles, reports, and guides, maintaining coherence over large text outputs. Limitations 1. Hardware Requirements: Requires high-memory GPUs or TPUs due to its large parameter size and long-context support. 2. Potential Bias in Responses: While designed to be neutral, outputs may still reflect biases present in training data. 3. Inconsistent Outputs in Creative Tasks: May produce variable results in storytelling and highly subjective topics. 4. Limited Real-World Awareness: Does not have access to real-time events beyond its training cutoff. 5. Error Propagation in Extended Outputs: Minor errors in early responses may affect overall coherence in long-form outputs. 6. Prompt Sensitivity: The effectiveness of responses may depend on how well the input prompt is structured. Open LLM Leaderboard Evaluation Results Detailed results can be found here! Summarized results can be found here! | Metric |Value (%)| |-------------------|--------:| |Average | 43.56| |IFEval (0-Shot) | 66.20| |BBH (3-Shot) | 59.92| |MATH Lvl 5 (4-Shot)| 34.74| |GPQA (0-shot) | 19.91| |MuSR (0-shot) | 28.49| |MMLU-PRO (5-shot) | 52.12|

Outfit-Cut is an adapter for black-forest-lab's FLUX.1-Kontext-dev, designed to extract outfits from images based on precisely specified subjects. The model was trained on 200 image pairs (100 start images and 100 end images). The adapter can be triggered with the following prompt: > [!note] [photo content], extract only the specified clothing item [full outfit, top wear, bottom wear, t-shirt, jacket, dress, etc.] from the image and place it over a clean, plain background. Present the result in a product photography style — well-lit, crisp, and professional — while preserving the garment’s original textures, colors, shapes, and fine details. > Prompt: [photo content], extract only the specified clothing item [t-shirt] from the image and place it over a clean, plain background. Present the result in a product photography style — well-lit, crisp, and professional — while preserving the garment’s original textures, colors, shapes, and fine details. > Prompt: [photo content], extract only the specified clothing item [top wear] from the image and place it over a clean, plain background. Present the result in a product photography style — well-lit, crisp, and professional — while preserving the garment’s original textures, colors, shapes, and fine details. > Prompt: [photo content], extract only the specified clothing item [t-shirt] from the image and place it over a clean, plain background. Present the result in a product photography style — well-lit, crisp, and professional — while preserving the garment’s original textures, colors, shapes, and fine details. > Prompt: [photo content], extract only the specified clothing item [t-shirt] from the image and place it over a clean, plain background. Present the result in a product photography style — well-lit, crisp, and professional — while preserving the garment’s original textures, colors, shapes, and fine details. > Prompt: [photo content], extract only the specified clothing item [t-shirt] from the image and place it over a clean, plain background. Present the result in a product photography style — well-lit, crisp, and professional — while preserving the garment’s original textures, colors, shapes, and fine details. > Prompt: [photo content], extract only the specified clothing item [t-shirt] from the image and place it over a clean, plain background. Present the result in a product photography style — well-lit, crisp, and professional — while preserving the garment’s original textures, colors, shapes, and fine details. > [!note] Note: This adapter works well for extracting top wear (t-shirts, shirts, jackets, hoodies). The model may perform sub-optimally in more challenging cases, such as full clothing extraction, poorly lit images, and other difficult scenarios. | Setting | Value | | ------------------------ | ------------------------ | | Module Type | Adapter | | Base Model | FLUX.1 Kontext Dev - fp8 | | Trigger Words | [photo content], extract only the specified clothing item [top wear, bottom wear] from the image and place it over a clean, plain background. Present the result in a product photography style — well-lit, crisp, and professional — while preserving the garment’s original textures, colors, shapes, and fine details. | | Image Processing Repeats | 50 | | Epochs | 25 | | Save Every N Epochs | 1 | Labeling: DeepCaption-VLA-7B(natural language & English) Total Images Used for Training : 200 Image Pairs (100 Start, 100 End) | Setting | Value | | --------------------------- | --------- | | Seed | - | | Clip Skip | - | | Text Encoder LR | 0.00001 | | UNet LR | 0.00005 | | LR Scheduler | constant | | Optimizer | AdamW8bit | | Network Dimension | 64 | | Network Alpha | 32 | | Gradient Accumulation Steps | - | | Setting | Value | | --------------- | ----- | | Shuffle Caption | - | | Keep N Tokens | - | | Setting | Value | | ------------------------- | ----- | | Noise Offset | 0.03 | | Multires Noise Discount | 0.1 | | Multires Noise Iterations | 10 | | Conv Dimension | - | | Conv Alpha | - | | Batch Size | - | | Steps | 2700 | | Sampler | euler | You should use `[photo content]` to trigger the image generation. You should use `extract only the specified clothing item [full outfit` to trigger the image generation. You should use `top wear` to trigger the image generation. You should use `bottom wear` to trigger the image generation. You should use `t-shirt` to trigger the image generation. You should use `jacket` to trigger the image generation. You should use `dress` to trigger the image generation. You should use `etc.] from the image and place it over a clean` to trigger the image generation. You should use `plain background. Present the result in a product photography style — well-lit` to trigger the image generation. You should use `crisp` to trigger the image generation. You should use `and professional — while preserving the garment’s original textures` to trigger the image generation. You should use `colors` to trigger the image generation. You should use `shapes` to trigger the image generation. You should use `and fine details.` to trigger the image generation.

> ReZero-v0.1-llama-3.2-3b-it-grpo-250404 is a research project focused on enhancing the search abilities of small language models by training them to develop robust search strategies rather than memorizing static data. The model, built on a Llama-3.2-3B backbone, interacts with multiple synthetic search engines that each have unique retrieval mechanisms, enabling it to refine queries iteratively and persist in finding exact answers using reinforcement learning. The repository provides setup instructions, including environment configuration and dependency installation, as well as scripts to train the model or regenerate synthetic training data. Demonstrations can be run through a Gradio interface, and the release includes comprehensive experiment logs on reward strategies and search quality. The model and associated resources are open-source and accessible to the research community, with further details on experiments and references provided in the documentation. | File name | Size | Quant Type | |-----------|------|------------| | llama-3.2-3b-it-grpo-250404.F32.gguf | 12.9 GB | F32 | | llama-3.2-3b-it-grpo-250404.BF16.gguf | 6.43 GB | BF16 | | llama-3.2-3b-it-grpo-250404.F16.gguf | 6.43 GB | F16 | | llama-3.2-3b-it-grpo-250404.Q80.gguf | 3.42 GB | Q80 | | llama-3.2-3b-it-grpo-250404.Q6K.gguf | 2.64 GB | Q6K | | llama-3.2-3b-it-grpo-250404.Q5KM.gguf | 2.32 GB | Q5KM | | llama-3.2-3b-it-grpo-250404.Q5KS.gguf | 2.27 GB | Q5KS | | llama-3.2-3b-it-grpo-250404.Q4KM.gguf | 2.02 GB | Q4KM | | llama-3.2-3b-it-grpo-250404.Q4KS.gguf | 1.93 GB | Q4KS | | llama-3.2-3b-it-grpo-250404.Q3KL.gguf | 1.82 GB | Q3KL | | llama-3.2-3b-it-grpo-250404.Q3KM.gguf | 1.69 GB | Q3KM | | llama-3.2-3b-it-grpo-250404.Q3KS.gguf | 1.54 GB | Q3KS | | llama-3.2-3b-it-grpo-250404.Q2K.gguf | 1.36 GB | Q2K | (sorted by size, not necessarily quality. IQ-quants are often preferable over similar sized non-IQ quants) Here is a handy graph by ikawrakow comparing some lower-quality quant types (lower is better):

> The Imgscope-OCR-2B-0527 model is a fine-tuned version of Qwen2-VL-2B-Instruct, specifically optimized for messy handwriting recognition, document OCR, realistic handwritten OCR, and math problem solving with LaTeX formatting. This model is trained on custom datasets for document and handwriting OCR tasks and integrates a conversational approach with strong visual and textual understanding for multi-modal applications. > [!note] Colab Demo : https://huggingface.co/prithivMLmods/Imgscope-OCR-2B-0527/blob/main/Imgscope%20OCR%202B%200527%20Demo/Imgscope-OCR-2B-0527.ipynb > [!note] Video Understanding Demo : https://huggingface.co/prithivMLmods/Imgscope-OCR-2B-0527/blob/main/Imgscope-OCR-2B-05270-Video-Understanding/Imgscope-OCR-2B-0527-Video-Understanding.ipynb SoTA Understanding of Images of Various Resolution & Ratio Imgscope-OCR-2B-0527 achieves state-of-the-art performance on visual understanding benchmarks such as MathVista, DocVQA, RealWorldQA, and MTVQA. Enhanced Handwriting OCR Specifically optimized for recognizing and interpreting realistic and messy handwriting with high accuracy. Ideal for digitizing handwritten documents and notes. Document OCR Fine-Tuning Fine-tuned with curated and realistic document OCR datasets, enabling accurate extraction of text from various structured and unstructured layouts. Understanding Videos of 20+ Minutes Capable of processing long videos for video-based question answering, transcription, and content generation. Device Control Agent Supports decision-making and control capabilities for integration with mobile devices, robots, and automation systems using visual-textual commands. Multilingual OCR Support In addition to English and Chinese, the model supports OCR in multiple languages including European languages, Japanese, Korean, Arabic, and Vietnamese. Fine-tuned for complex and hard-to-read handwritten inputs using real-world handwriting datasets. Accurately extracts text from structured documents, including scanned pages, forms, and academic papers. Combines vision-language capabilities for tasks like captioning, answering image-based queries, and understanding image+text prompts. Converts mathematical expressions and problem-solving steps into LaTeX format. Supports dialogue-based reasoning, retaining context for follow-up questions. Accepts inputs from videos, images, or combined media with text, and generates relevant output accordingly. Handwritten and printed document digitization OCR pipelines for educational institutions and businesses Academic and scientific content parsing, especially math-heavy documents Assistive tools for visually impaired users Robotic and mobile automation agents interpreting screen or camera data Multilingual OCR processing for document translation or archiving

> `Forest-Fire-Detection` is a vision-language encoder model fine-tuned from `google/siglip2-base-patch16-512` for multi-class image classification. It is trained to detect whether an image contains fire, smoke, or a normal (non-fire) scene. The model uses the `SiglipForImageClassification` architecture. > [!note] SigLIP 2: Multilingual Vision-Language Encoders with Improved Semantic Understanding, Localization, and Dense Features : https://arxiv.org/pdf/2502.14786 Wildfire Monitoring – Rapid identification of forest fire and smoke zones. Environmental Protection – Surveillance of forest areas for early fire warning. Disaster Management – Support in emergency response and evacuation decisions. Smart Surveillance – Integrate with drones or camera feeds for automated fire detection. Research and Analysis – Analyze visual datasets for fire-prone region identification.

> [!note] Qwen-Image-HeadshotX is a super-realistic headshot adapter for Qwen-Image, an image generation model by Qwen. It is an advanced LoRA adaptation of the Qwen-Image model and an upgraded version of Qwen-Image-Studio-Realism, offering more precise portrait rendering with a strong focus on realism. The model was trained on diverse face types from across the world, labeled with florence2-en and caption-optimized using DeepCaption-VLA-7B. Total Images Used for Training: 55 RAW images \[11(types) × 5 different face types: Asian, Hispanic, Caucasian, Latina, Middle Eastern, etc.]. | Parameter | Value | Parameter | Value | |---------------------------|--------|---------------------------|--------| | LR Scheduler | constant | Noise Offset | 0.03 | | Optimizer | AdamW | Multires Noise Discount | 0.1 | | Network Dim | 64 | Multires Noise Iterations | 10 | | Network Alpha | 32 | Repeat & Steps | 25 & 4000 | | Epoch | 30 | Save Every N Epochs | 1 | Labeling: florence2-en(natural language & English) + 🔥 Optimized with Long-Caption VLA Multimodal : https://huggingface.co/prithivMLmods/DeepCaption-VLA-7B Total Images Used for Training: 55 RAW [11(types)×5 different face types (Asian, Hispanic, Caucasian, Latina, Middle Eastern, etc.)] | Source | Link | |--------------|-------------------------------------| | Playground | playground.com | | ArtStation | artstation.com | | 4K Wallpapers| 4kwallpapers.com | | Dimensions | Aspect Ratio | Recommendation | |-----------------|------------------|---------------------------| | 1472 x 1140 | 4:3 (approx.) | Best | | 1024 x 1024 | 1:1 | Default | - Recommended Inference Steps: 45-50 (approx. ~ `100 Seconds Inference`) You should use `face headshot` to trigger the image generation.

> Demeter-LongCoT-Qwen3-1.7B is a reasoning-focused model fine-tuned on Qwen/Qwen3-1.7B using the Demeter-LongCoT-400K dataset. It is designed for math and code chain-of-thought reasoning, blending symbolic precision, scientific logic, and structured output fluency—making it an effective tool for developers, educators, and researchers seeking reliable step-by-step reasoning. | File Name | Quant Type | File Size | | - | - | - | | Demeter-LongCoT-Qwen-1.7B.BF16.gguf | BF16 | 3.45 GB | | Demeter-LongCoT-Qwen-1.7B.F16.gguf | F16 | 3.45 GB | | Demeter-LongCoT-Qwen-1.7B.F32.gguf | F32 | 6.89 GB | | Demeter-LongCoT-Qwen-1.7B.Q2K.gguf | Q2K | 778 MB | | Demeter-LongCoT-Qwen-1.7B.Q3KL.gguf | Q3KL | 1 GB | | Demeter-LongCoT-Qwen-1.7B.Q3KM.gguf | Q3KM | 940 MB | | Demeter-LongCoT-Qwen-1.7B.Q3KS.gguf | Q3KS | 867 MB | | Demeter-LongCoT-Qwen-1.7B.Q40.gguf | Q40 | 1.05 GB | | Demeter-LongCoT-Qwen-1.7B.Q41.gguf | Q41 | 1.14 GB | | Demeter-LongCoT-Qwen-1.7B.Q4K.gguf | Q4K | 1.11 GB | | Demeter-LongCoT-Qwen-1.7B.Q4KM.gguf | Q4KM | 1.11 GB | | Demeter-LongCoT-Qwen-1.7B.Q4KS.gguf | Q4KS | 1.06 GB | | Demeter-LongCoT-Qwen-1.7B.Q50.gguf | Q50 | 1.23 GB | | Demeter-LongCoT-Qwen-1.7B.Q51.gguf | Q51 | 1.32 GB | | Demeter-LongCoT-Qwen-1.7B.Q5K.gguf | Q5K | 1.26 GB | | Demeter-LongCoT-Qwen-1.7B.Q5KM.gguf | Q5KM | 1.26 GB | | Demeter-LongCoT-Qwen-1.7B.Q5KS.gguf | Q5KS | 1.23 GB | | Demeter-LongCoT-Qwen-1.7B.Q6K.gguf | Q6K | 1.42 GB | | Demeter-LongCoT-Qwen-1.7B.Q80.gguf | Q80 | 1.83 GB | (sorted by size, not necessarily quality. IQ-quants are often preferable over similar sized non-IQ quants) Here is a handy graph by ikawrakow comparing some lower-quality quant types (lower is better):

> Leporis-Qwen3-Radiation-1.7B is a reasoning-focused model fine-tuned on Qwen for Abliterated Reasoning and polished token probabilities, enhancing balanced multilingual generation across mathematics and general-purpose reasoning. It specializes in event-driven logic, structured analysis, and precise probabilistic modeling—making it an ideal tool for researchers, educators, and developers working with uncertainty and structured reasoning. | File Name | Quant Type | File Size | | - | - | - | | Leporis-Qwen3-Radiation-1.7B.BF16.gguf | BF16 | 3.45 GB | | Leporis-Qwen3-Radiation-1.7B.F16.gguf | F16 | 3.45 GB | | Leporis-Qwen3-Radiation-1.7B.F32.gguf | F32 | 6.89 GB | | Leporis-Qwen3-Radiation-1.7B.Q2K.gguf | Q2K | 778 MB | | Leporis-Qwen3-Radiation-1.7B.Q3KL.gguf | Q3KL | 1 GB | | Leporis-Qwen3-Radiation-1.7B.Q3KM.gguf | Q3KM | 940 MB | | Leporis-Qwen3-Radiation-1.7B.Q3KS.gguf | Q3KS | 867 MB | | Leporis-Qwen3-Radiation-1.7B.Q40.gguf | Q40 | 1.05 GB | | Leporis-Qwen3-Radiation-1.7B.Q41.gguf | Q41 | 1.14 GB | | Leporis-Qwen3-Radiation-1.7B.Q4K.gguf | Q4K | 1.11 GB | | Leporis-Qwen3-Radiation-1.7B.Q4KM.gguf | Q4KM | 1.11 GB | | Leporis-Qwen3-Radiation-1.7B.Q4KS.gguf | Q4KS | 1.06 GB | | Leporis-Qwen3-Radiation-1.7B.Q50.gguf | Q50 | 1.23 GB | | Leporis-Qwen3-Radiation-1.7B.Q51.gguf | Q51 | 1.32 GB | | Leporis-Qwen3-Radiation-1.7B.Q5K.gguf | Q5K | 1.26 GB | | Leporis-Qwen3-Radiation-1.7B.Q5KM.gguf | Q5KM | 1.26 GB | | Leporis-Qwen3-Radiation-1.7B.Q5KS.gguf | Q5KS | 1.23 GB | | Leporis-Qwen3-Radiation-1.7B.Q6K.gguf | Q6K | 1.42 GB | | Leporis-Qwen3-Radiation-1.7B.Q80.gguf | Q80 | 1.83 GB | (sorted by size, not necessarily quality. IQ-quants are often preferable over similar sized non-IQ quants) Here is a handy graph by ikawrakow comparing some lower-quality quant types (lower is better):

> DynaGuard is a dynamic guardrail model framework that enables user-defined content moderation policies for generative AI, providing real-time, configurable controls on both model inputs and outputs. It employs ultralightweight, efficient models optimized for on-device or cloud deployments, allowing organizations to protect against risks like data leakage, PII exposure, toxicity, jailbreaking, and prompt injection with sub-50ms latency and minimal compute overhead. Users can create custom guardrails in natural language, tailoring moderation behaviors, targeted content categories, and downstream actions, making DynaGuard ideal for enterprise AI security, compliance, and safe application development across diverse hardware platforms. | File Name | Quant Type | File Size | | - | - | - | | DynaGuard-1.7B.BF16.gguf | BF16 | 3.45 GB | | DynaGuard-1.7B.F16.gguf | F16 | 3.45 GB | | DynaGuard-1.7B.F32.gguf | F32 | 6.89 GB | | DynaGuard-1.7B.Q2K.gguf | Q2K | 778 MB | | DynaGuard-1.7B.Q3KL.gguf | Q3KL | 1 GB | | DynaGuard-1.7B.Q3KM.gguf | Q3KM | 940 MB | | DynaGuard-1.7B.Q3KS.gguf | Q3KS | 867 MB | | DynaGuard-1.7B.Q40.gguf | Q40 | 1.05 GB | | DynaGuard-1.7B.Q41.gguf | Q41 | 1.14 GB | | DynaGuard-1.7B.Q4K.gguf | Q4K | 1.11 GB | | DynaGuard-1.7B.Q4KM.gguf | Q4KM | 1.11 GB | | DynaGuard-1.7B.Q4KS.gguf | Q4KS | 1.06 GB | | DynaGuard-1.7B.Q50.gguf | Q50 | 1.23 GB | | DynaGuard-1.7B.Q51.gguf | Q51 | 1.32 GB | | DynaGuard-1.7B.Q5K.gguf | Q5K | 1.26 GB | | DynaGuard-1.7B.Q5KM.gguf | Q5KM | 1.26 GB | | DynaGuard-1.7B.Q5KS.gguf | Q5KS | 1.23 GB | | DynaGuard-1.7B.Q6K.gguf | Q6K | 1.42 GB | | DynaGuard-1.7B.Q80.gguf | Q80 | 1.83 GB | | File Name | Quant Type | File Size | | - | - | - | | DynaGuard-4B.BF16.gguf | BF16 | 8.05 GB | | DynaGuard-4B.F16.gguf | F16 | 8.05 GB | | DynaGuard-4B.F32.gguf | F32 | 16.1 GB | | DynaGuard-4B.Q2K.gguf | Q2K | 1.67 GB | | DynaGuard-4B.Q3KL.gguf | Q3KL | 2.24 GB | | DynaGuard-4B.Q3KM.gguf | Q3KM | 2.08 GB | | DynaGuard-4B.Q3KS.gguf | Q3KS | 1.89 GB | | DynaGuard-4B.Q4KM.gguf | Q4KM | 2.5 GB | | DynaGuard-4B.Q4KS.gguf | Q4KS | 2.38 GB | | DynaGuard-4B.Q5KM.gguf | Q5KM | 2.89 GB | | DynaGuard-4B.Q5KS.gguf | Q5KS | 2.82 GB | | DynaGuard-4B.Q6K.gguf | Q6K | 3.31 GB | | DynaGuard-4B.Q80.gguf | Q80 | 4.28 GB | (sorted by size, not necessarily quality. IQ-quants are often preferable over similar sized non-IQ quants) Here is a handy graph by ikawrakow comparing some lower-quality quant types (lower is better):

Deepfake-Detection-Exp-02-22 is a minimalist, high-quality dataset trained on a ViT-based model for image classification, distinguishing between deepfake and real images. The model is based on Google's `google/vit-base-patch32-224-in21k`. Limitations 1. Generalization Issues – The model may not perform well on deepfake images generated by unseen or novel deepfake techniques. 2. Dataset Bias – The training data might not cover all variations of real and fake images, leading to biased predictions. 3. Resolution Constraints – Since the model is based on `vit-base-patch32-224-in21k`, it is optimized for 224x224 image resolution, which may limit its effectiveness on high-resolution images. 4. Adversarial Vulnerabilities – The model may be susceptible to adversarial attacks designed to fool vision transformers. 5. False Positives & False Negatives – The model may occasionally misclassify real images as deepfake and vice versa, requiring human validation in critical applications. Intended Use 1. Deepfake Detection – Designed for identifying deepfake images in media, social platforms, and forensic analysis. 2. Research & Development – Useful for researchers studying deepfake detection and improving ViT-based classification models. 3. Content Moderation – Can be integrated into platforms to detect and flag manipulated images. 4. Security & Forensics – Assists in cybersecurity applications where verifying the authenticity of images is crucial. 5. Educational Purposes – Can be used in training AI practitioners and students in the field of computer vision and deepfake detection.

Evac-Opus-14B-Exp [abliterated] is an advanced language model based on the Qwen 2.5 14B modality architecture, designed to enhance reasoning, explanation, and conversational capabilities. This model is optimized for general-purpose tasks, excelling in contextual understanding, logical deduction, and multi-step problem-solving. It has been fine-tuned using a long chain-of-thought reasoning model and specialized datasets to improve comprehension, structured responses, and conversational intelligence. Key Improvements 1. Enhanced General Knowledge: The model provides broad knowledge across various domains, improving capabilities in answering questions accurately and generating coherent responses. 2. Improved Instruction Following: Significant advancements in understanding and following complex instructions, generating structured responses, and maintaining coherence over extended interactions. 3. Versatile Adaptability: More resilient to diverse prompts, enhancing its ability to handle a wide range of topics and conversation styles, including open-ended and structured inquiries. 4. Long-Context Support: Supports up to 128K tokens for input context and can generate up to 8K tokens in a single output, making it ideal for detailed responses. 5. Multilingual Proficiency: Supports over 29 languages, including English, Chinese, French, Spanish, Portuguese, German, Italian, Russian, Japanese, Korean, Vietnamese, Thai, Arabic, and more. Here is a code snippet with `applychattemplate` to show you how to load the tokenizer and model and generate content: Intended Use 1. General-Purpose Reasoning: Designed for broad applicability, assisting with logical reasoning, answering diverse questions, and solving general knowledge problems. 2. Educational and Informational Assistance: Suitable for providing explanations, summaries, and research-based responses for students, educators, and general users. 3. Conversational AI and Chatbots: Ideal for building intelligent conversational agents that require contextual understanding and dynamic response generation. 4. Multilingual Applications: Supports global communication, translations, and multilingual content generation. 5. Structured Data Processing: Capable of analyzing and generating structured outputs, such as tables and JSON, useful for data science and automation. 6. Long-Form Content Generation: Can generate extended responses, including articles, reports, and guides, maintaining coherence over large text outputs. Limitations 1. Hardware Requirements: Requires high-memory GPUs or TPUs due to its large parameter size and long-context support. 2. Potential Bias in Responses: While designed to be neutral, outputs may still reflect biases present in training data. 3. Inconsistent Outputs in Creative Tasks: May produce variable results in storytelling and highly subjective topics. 4. Limited Real-World Awareness: Does not have access to real-time events beyond its training cutoff. 5. Error Propagation in Extended Outputs: Minor errors in early responses may affect overall coherence in long-form outputs. 6. Prompt Sensitivity: The effectiveness of responses may depend on how well the input prompt is structured. Open LLM Leaderboard Evaluation Results Detailed results can be found here! Summarized results can be found here! | Metric |Value (%)| |-------------------|--------:| |Average | 39.32| |IFEval (0-Shot) | 59.16| |BBH (3-Shot) | 49.58| |MATH Lvl 5 (4-Shot)| 42.15| |GPQA (0-shot) | 18.46| |MuSR (0-shot) | 18.63| |MMLU-PRO (5-shot) | 47.96|

Llama-3.1 is a collection of multilingual large language models (LLMs) that includes pretrained and instruction-tuned generative models in various sizes. The Llama-3.1-5B-Instruct model is part of the series optimized for multilingual dialogue use cases, offering powerful conversational abilities and outperforming many open-source and closed chat models on key industry benchmarks. - Size: 5B parameters - Model Architecture: Llama-3.1 is an auto-regressive language model using an optimized transformer architecture. - Training: The model is fine-tuned using Supervised Fine-Tuning (SFT) and Reinforcement Learning with Human Feedback (RLHF) to align with human preferences, ensuring helpfulness, safety, and natural conversations. The Llama-3.1-5B-Instruct model is optimized for multilingual text generation and excels in a variety of dialog-based use cases. It is designed to handle a wide array of tasks, including question answering, translation, and instruction following. - Ensure you have PyTorch installed with support for `bfloat16`: Below is an example of how to use the Llama-3.1-5B-Instruct model for conversational inference: - Model Type: Instruction-Tuned Large Language Model (LLM) - Training: Trained using supervised fine-tuning and reinforcement learning with human feedback. - Supported Tasks: Dialogue generation, question answering, translation, and other text-based tasks. The Llama-3.1-5B-Instruct model outperforms many existing models on several benchmarks, making it a reliable choice for conversational AI tasks in multilingual environments. - This model is optimized for safety and helpfulness, ensuring a positive user experience. - The torchdtype is set to `bfloat16` to optimize memory usage and performance. ---

> Qwen3-4B-MegaScience is a large language model fine-tuned on the MegaScience dataset, specifically designed for advanced scientific reasoning, and built on top of Qwen3-4B-Base; it leverages a meticulously curated set of 1.25 million high-quality scientific questions and answers sourced from university-level textbooks and various open datasets, covering seven scientific disciplines and evaluated across 15 benchmarks, demonstrating superior reasoning ability and training efficiency compared to existing open-source science models; the model supports seamless integration via the Hugging Face transformers library, operates efficiently with bfloat16 precision, and comes with an open-source dataset, evaluation pipeline, and reproducibility code, facilitating research and applications in scientific AI reasoning, with full resources, paper, and code available via the MegaScience official website and GitHub repository. | File Name | Size | Quant Type | |-----------|------|------------| | Qwen3-4B-MegaScience.BF16.gguf | 8.05 GB | BF16 | | Qwen3-4B-MegaScience.F16.gguf | 8.05 GB | F16 | | Qwen3-4B-MegaScience.F32.gguf | 16.1 GB | F32 | | Qwen3-4B-MegaScience.Q3KL.gguf | 2.24 GB | Q3KL | | Qwen3-4B-MegaScience.Q3KS.gguf | 1.89 GB | Q3KS | | Qwen3-4B-MegaScience.Q4KM.gguf | 2.5 GB | Q4KM | | Qwen3-4B-MegaScience.Q4KS.gguf | 2.38 GB | Q4KS | | Qwen3-4B-MegaScience.Q5KM.gguf | 2.89 GB | Q5KM | | Qwen3-4B-MegaScience.Q5KS.gguf | 2.82 GB | Q5KS | | Qwen3-4B-MegaScience.Q6K.gguf | 3.31 GB | Q6K | | Qwen3-4B-MegaScience.Q80.gguf | 4.28 GB | Q80 | (sorted by size, not necessarily quality. IQ-quants are often preferable over similar sized non-IQ quants) Here is a handy graph by ikawrakow comparing some lower-quality quant types (lower is better):

FastThink-0.5B-Tiny is a reasoning-focused model based on Qwen2.5. We have released a range of base language models and instruction-tuned language models, spanning from 0.5 billion to 72 billion parameters. Qwen2.5 introduces the following improvements over Qwen2: - Significantly enhanced knowledge and greatly improved capabilities in coding and mathematics, thanks to specialized expert models in these domains. - Major improvements in instruction following, generating long texts (over 8K tokens), understanding structured data (e.g., tables), and generating structured outputs, especially JSON. It is more resilient to diverse system prompts, enhancing role-play implementation and condition-setting for chatbots. - Long-context support for up to 128K tokens and the ability to generate outputs up to 8K tokens. - Multilingual support for over 29 languages, including Chinese, English, French, Spanish, Portuguese, German, Italian, Russian, Japanese, Korean, Vietnamese, Thai, Arabic, and more. Architecture: Transformers with RoPE, SwiGLU, RMSNorm, Attention QKV bias, and tied word embeddings. Here provides a code snippet with `applychattemplate` to show you how to load the tokenizer and model and how to generate contents. This script is designed to load, process, and combine multiple datasets into a single, standardized format suitable for training conversational AI models. The script uses the `datasets` library to load and manipulate the datasets, and the `chattemplates` library to standardize the conversation format. Intended Use 1. Reasoning Tasks: FastThink-0.5B-Tiny is optimized for reasoning-focused applications, such as logical problem-solving, decision-making, and analytical workflows. 2. Instruction Following: Ideal for scenarios where precise adherence to instructions is required, including generating structured outputs like JSON or tables. 3. Multilingual Support: Suitable for use in multilingual environments, supporting over 29 languages, making it versatile for global applications. 4. Coding and Mathematics: Highly effective in tasks involving coding, debugging, or solving mathematical problems, leveraging expert domain knowledge. 5. Role-play Scenarios: Can simulate conversational agents or personas for role-playing, enhancing chatbot and virtual assistant implementations. 6. Long-form Content Creation: Designed to generate and manage long-form text (up to 8K tokens) while maintaining context, making it ideal for tasks like report writing or storytelling. 7. Understanding and Processing Structured Data: Efficient at interpreting and working with structured data, such as tables or hierarchical formats. 8. Low-Resource Applications: With a smaller parameter size (0.5B), it is well-suited for applications with limited computational resources or edge deployment. Limitations 1. Limited Model Size: As a 0.5B-parameter model, its reasoning and comprehension capabilities are less advanced compared to larger models, particularly for highly complex tasks. 2. Contextual Limitations: Although it supports a context length of up to 128K tokens, its ability to effectively utilize such a long context may vary, particularly in tasks requiring intricate cross-referencing of earlier inputs. 3. Accuracy in Domain-Specific Tasks: While capable in coding and mathematics, it may struggle with highly specialized or esoteric domain knowledge compared to models fine-tuned specifically for those areas. 4. Ambiguity Handling: May misinterpret vague or poorly structured prompts, leading to less accurate or unintended results. 5. Long-Context Tradeoffs: Generating or processing very long outputs (e.g., close to the 8K token limit) could result in decreased coherence or relevance toward the end. 6. Multilingual Performance: Although it supports 29 languages, its proficiency and fluency may vary across languages, with some underrepresented languages possibly seeing reduced performance. 7. Resource-Intensive for Long Contexts: Using its long-context capabilities (128K tokens) can be computationally demanding, requiring significant memory and processing power. 8. Dependence on Fine-Tuning: For highly specialized tasks or domains, additional fine-tuning may be necessary to achieve optimal performance.

> The DeepCaption-VLA-7B model is a fine-tuned version of Qwen2.5-VL-7B-Instruct, tailored for Image Captioning and Vision Language Attribution. This variant is designed to generate precise, highly descriptive captions with a focus on defining visual properties, object attributes, and scene details across a wide spectrum of images and aspect ratios. [](https://github.com/PRITHIVSAKTHIUR/Multimodal-Outpost-Notebooks/blob/main/DeepCaption-VLA-7B%5B4bit%20-%20notebook%20demo%5D/DeepCaption-VLA-7B.ipynb) 1. Vision Language Attribution (VLA): Specially fine-tuned to attribute and define visual properties of objects, scenes, and environments. 2. Detailed Object Definitions: Generates captions with rich attribute descriptions, making outputs more precise than generic captioners. 3. High-Fidelity Descriptions: Handles general, artistic, technical, abstract, and low-context images with descriptive depth. 4. Robust Across Aspect Ratios: Accurately captions images regardless of format—wide, tall, square, or irregular. 5. Variational Detail Control: Supports both concise summaries and fine-grained attributions depending on prompt structure. 6. Foundation on Qwen2.5-VL Architecture: Leverages Qwen2.5-VL-7B’s multimodal reasoning for visual comprehension and instruction-following. 7. Multilingual Capability: Default in English, but adaptable for multilingual captioning through prompt engineering. This model was fine-tuned with a curated mix of datasets focused on caption richness and object-attribute alignment: prithivMLmods/blip3o-caption-mini-arrow prithivMLmods/Caption3o-Opt-v3 prithivMLmods/Caption3o-Opt-v2 Multimodal-Fatima/Caltech101\not\background\test\facebook\opt\2.7b\Attributes\Caption\ns\5647 Private/unlisted datasets for domain-specific image captioning tasks. The training objective emphasized Vision Language Attribution: defining image properties, attributes, and objects with clarity, while preserving descriptive fluency. > [!note] General Query: Caption the image precisely. [](https://colab.research.google.com/#fileId=https%3A//huggingface.co/prithivMLmods/DeepCaption-VLA-7B/blob/main/DeepCaption-VLA-7B%5B4bit%20-%20notebook%20demo%5D/DeepCaption-VLA-7B.ipynb) Generating attribute-rich image captions for research, dataset creation, and AI training. Vision-language attribution for object detection, scene understanding, and dataset annotation. Supporting creative, artistic, and technical applications requiring detailed descriptions. Captioning across varied aspect ratios, unusual visual styles, and non-standard datasets. May over-attribute or infer properties not explicitly visible in ambiguous images. Outputs can vary in tone depending on prompt phrasing. Not intended for filtered captioning tasks (explicit or sensitive content may appear). Accuracy may degrade on synthetic or highly abstract visual domains.