SVECTOR-CORPORATION

We are excited to present .dotcode-1-mini, a compact and efficient language model developed by SVECTOR. This model represents our commitment to building accessible, high-performance AI solutions that empower developers and researchers. - Efficiency: Optimized architecture for fast inference and reduced computational requirements - Versatility: Strong performance across diverse text generation and code-related tasks - Accessibility: Open-source model available to the community under Apache 2.0 license Balanced approach to capability and resource efficiency. - Type: Causal language model (LLaMA-based architecture) - License: Apache 2.0 - Context Length: 32K To use .dotcode-1-mini, ensure you have the latest versions of `transformers` and `accelerate` installed: Here's a simple example demonstrating how to load and use the model: - Code Generation: Writing functions, scripts, and complete programs - Text Completion: Intelligent continuation of text and code - Problem Solving: Logical reasoning and algorithmic thinking - Documentation: Generating comments, docstrings, and technical explanations - General Text Generation: Creative writing, summaries, and content creation .dotcode-1-mini has been designed to provide strong performance while maintaining a compact model size. Detailed benchmarks and evaluation results will be shared as they become available. Built on the LLaMA architecture, .dotcode-1-mini incorporates optimizations specifically tailored for: - Efficient token processing - Reduced memory footprint - Fast inference speeds - Balanced precision and performance .dotcode-1-mini was trained on a diverse corpus including: - High-quality code repositories - Technical documentation - General text data - Curated datasets for improved reasoning Detailed training methodology and data composition will be documented in future releases. As with any language model, .dotcode-1-mini has certain limitations: - May generate incorrect or outdated information - Performance varies based on prompt quality and task complexity - Not specifically fine-tuned for specialized domains without additional training - Should be used with appropriate safeguards in production environments SVECTOR is committed to responsible AI development. Users should: - Review outputs for accuracy and appropriateness - Implement content filtering for sensitive applications - Avoid using the model for harmful or malicious purposes - Respect copyright and intellectual property when generating code This model is released under the Apache License 2.0. See the LICENSE file for complete details.

Theta-35 is the advanced reasoning model in the Theta series by SVECTOR. Compared with conventional instruction-tuned models, Theta-35, which specializes in complex thinking and reasoning, achieves significantly enhanced performance in downstream tasks, particularly for challenging problems requiring deep logical analysis and multistep reasoning. This repo contains the Theta-35 model, which has the following features: - Training Stage: Pretraining & Post-training (Supervised Finetuning and Reinforcement Learning) - Architecture: Transformers with RoPE, SwiGLU, RMSNorm, and Attention QKV bias - Number of Parameters: 33B - Number of Parameters (Non-Embedding): 33B - Number of Layers: 64 - Number of Attention Heads (GQA): 40 for Q and 8 for KV - Context Length: Full 32k tokens - Sliding Window: 32,768 tokens Note: For the best experience, please review the usage guidelines before deploying Theta models. For more details, please refer to our documentation. Theta-35 requires the latest version of Hugging Face `transformers`. We advise you to use version 4.43.1 or newer. With older versions of transformers, you may encounter the following error: Here is a code snippet showing how to load the tokenizer and model, and how to generate content: To achieve optimal performance with Theta-35, we recommend the following settings: 1. Enforce Thoughtful Output: Ensure the model starts with "\ \n" to promote step-by-step thinking, which enhances output quality. If you use `applychattemplate` and set `addgenerationprompt=True`, this is automatically implemented. 2. Sampling Parameters: - Use Temperature=0.6 and TopP=0.95 instead of Greedy decoding to avoid repetitions. - Use TopK between 20 and 40 to filter out rare token occurrences while maintaining diversity. 3. Standardize Output Format: We recommend using prompts to standardize model outputs when benchmarking. - Math Problems: Include "Please reason step by step, and put your final answer within \boxed{}." in the prompt. - Multiple-Choice Questions: Add "Please show your choice in the `answer` field with only the choice letter, e.g.,`\"answer\": \"C\"`." to the prompt. 4. Handle Long Inputs: For inputs exceeding 32,768 tokens, enable sliding window attention to improve the model's ability to process long sequences efficiently. For supported frameworks, you could add the following to `config.json` to enable extended context handling: Theta-35 demonstrates exceptional performance across various reasoning tasks, including: - Mathematical reasoning - Logical deduction - Multi-step problem solving - Code understanding and generation - Scientific reasoning Detailed evaluation results are reported in our documentation. If you find our work helpful, feel free to give us a cite.

Fal-2-500M is a compact vision-language model designed for image understanding and captioning tasks. Built on the Qwen2 architecture with an efficient vision encoder, it provides high-quality image descriptions with fast inference. - Model Size: 500M parameters We introduce a hybrid vision encoder designed to output fewer tokens and significantly reduce encoding time for high-resolution images. - Efficient Token Generation: 256 tokens at 1024×1024 resolution (16× fewer than ViT) - State-of-the-Art Performance: Competitive accuracy with superior efficiency - Primary Use: Image captioning and visual question answering

We’re proud to unveil Akshara-8B, our cutting-edge AI fleet built exclusively for India’s diverse linguistic landscape. Akshara is designed to seamlessly understand and generate text in multiple Indian languages, making AI more accessible, powerful, and tailored to our nation’s needs. Akshara-8B is a highly optimized distilled version of SVECTOR’s flagship large-scale AI model (Akshara). While it retains the core intelligence and multilingual capabilities of its parent model, Akshara-8B is specifically designed for efficiency, speed, and accessibility. It leverages advanced distillation techniques to provide powerful AI performance while being lightweight and scalable. Akshara-8B embodies SVECTOR’s commitment to bringing cutting-edge AI to India, ensuring robust support for India’s diverse languages and applications. 🚀 Akshara can fluently understand and generate content in: ✅ Hindi ✅ Gujarati ✅ Marathi ✅ Tamil ✅ Telugu ✅ Kannada ✅ Punjabi ✅ English 🔥 Why Akshara? 🔹 Made in India, for India & Global 🇮🇳 🔹 Optimized for speed and efficiency ⚡ 🔹 Seamless multilingual processing 🗣️ 🔹 Balanced accuracy and creativity 🎨 🔹 Lightweight and scalable for real-world applications 🚀 💬 Multi-turn Conversation Support Akshara supports multi-turn, dynamic conversations across languages. 🌟 Akshara: Built for the Future of AI in India By embracing India’s linguistic diversity, Akshara represents a major step toward bridging the AI gap in our country. Whether it's education, research, customer service, content creation, or smart automation, Akshara is here to revolutionize multilingual AI interactions. Join us as we shape the future of AI for India! 🇮🇳🚀 ```bibtex @misc{SVECTOR2025Akshara, title = {Akshara: A Multilingual AI Model for India}, author = {SVECTOR}, year = {2025}, url = {https://svector.co.in}, note = {Developed by SVECTOR CORPORATION for multilingual AI Model}, }

We are thrilled to introduce Continue-1-OSS, an advanced text generation model developed by SVECTOR, built on the Continue-1 architecture optimized for high-quality text generation, instruction following, and long-context understanding. - Superior Instruction Following: Accurately follows complex, multi-step instructions - Long Context: Robust handling of up to 128K+ tokens - Natural Conversations: Human-like dialogue with strong reasoning capabilities - Tool Integration: Built-in support for function calling and external tool use - Open Source: Fully accessible under Apache 2.0 license for research and commercial use This model combines the power of transformer architecture with advanced training techniques to deliver exceptional performance across a wide range of natural language tasks. - Base Architecture: Continue1ForCausalLM (transformer decoder) - Model Type: continueoss - Parameters: 3 Billion - Context Length: 131,072 tokens - Vocabulary Size: 128,256 tokens - Hidden Size: 3072 - Number of Layers: 28 - Attention Heads: 24 - License: Apache 2.0 To use Continue-1-OSS, install the required dependencies: For high-performance inference with faster generation: Default System Prompt: "You are Continue-1-OSS, an advanced AI assistant developed by SVECTOR. You are designed to be helpful, harmless, and honest." Continue-1-OSS supports function calling for tool integration: - Conversational AI: Build chatbots and virtual assistants with natural dialogue - Content Generation: Generate articles, stories, and creative content - Code Assistance: Help with coding tasks, debugging, and code explanations - Question Answering: Answer questions based on context with high accuracy - Summarization: Condense long documents into concise summaries - Data Extraction: Extract structured data from unstructured text - Tool Integration: Call functions and use external tools intelligently - Education: Create educational content and tutoring assistance - Customer Service: Automated support with natural language understanding - Quality: State-of-the-art instruction following and text generation - Speed: Fast inference with vLLM optimization - Memory: ~7GB GPU RAM (BF16), ~14GB (FP32) - Context: Handles up to 128K tokens effectively - Efficiency: Competitive with much larger models on many tasks Continue-1-OSS uses a custom architecture based on the transformer decoder: - Architecture Class: `Continue1ForCausalLM` - Config Class: `Continue1Config` - Hidden Size: 3072 - Num Layers: 28 - Num Attention Heads: 24 - Intermediate Size: 8192 - Vocab Size: 128,256 - Max Position Embeddings: 131,072 The model uses RoPE (Rotary Position Embeddings) for positional encoding and supports extended context through position interpolation. Continue-1-OSS was developed using: - High-quality instruction datasets covering diverse tasks - Conversational and reasoning data for improved dialogue - Code and technical content for developer assistance - Multi-turn dialogue for contextual understanding Training utilized: - Advanced optimization techniques - Careful hyperparameter tuning - Quality filtering and data curation - Evaluation on diverse benchmarks As with any language model, Continue-1-OSS has certain limitations: - Knowledge Cutoff: Training data is limited to information available up to December 2023 - Factual Accuracy: May occasionally generate incorrect or outdated information - Specialized Domains: Performance may vary on highly specialized technical knowledge - Long Context: Very long contexts (>64K tokens) may impact generation quality - Languages: Primarily optimized for English; other languages have limited support - Reasoning: Complex multi-step reasoning may require careful prompting - Compute: Requires GPU for optimal performance (CPU is significantly slower) SVECTOR is committed to responsible AI development. Users should: - Transparency: Disclose when content is AI-generated - Verification: Always fact-check important information generated by the model - Bias Awareness: Be aware the model may reflect biases present in training data - Privacy: Do not input personal or sensitive information without proper safeguards - Safety: Implement content filtering and guardrails for production applications - Responsible Use: Do not use for illegal purposes, misinformation, or harmful content - Attribution: Credit the model when used in public projects or research 1. Temperature Settings: - 0.0-0.3 for factual/deterministic tasks - 0.7-0.9 for creative tasks 2. Context Management: - Model supports 128K tokens but consider truncating for faster inference - Use sliding window for very long documents 3. Batch Processing: - Use vLLM for efficient batched inference in production - Group similar-length prompts together This model is released under the Apache License 2.0. You are free to use, modify, and distribute this model for both commercial and non-commercial purposes. See the LICENSE file for complete details.

We are thrilled to introduce Continue-TTS, a fine-tuned text-to-speech model based on the Continue-1-OSS architecture, developed by SVECTOR. This model is specifically trained for high-quality speech synthesis and delivers exceptional voice generation capabilities. - Natural Speech: Human-like intonation, emotion, and rhythm that rivals commercial solutions - 8 Unique Voices: Diverse voice options with distinct personalities and characteristics - Real-time Generation: Low-latency streaming for interactive applications (~200ms) - Emotional Expression: Built-in support for laughter, sighs, gasps, and other natural emotions - Open Source: Fully accessible under Apache 2.0 license for research and commercial use This model is based on the Continue-1-OSS architecture and combines the power of large language models with neural audio codecs to generate exceptionally natural speech from text. - Base Architecture: Continue-1-OSS - Type: Text-to-Speech (TTS) Model - Parameters: 3 Billion - Audio Codec: SNAC (24kHz) - Context Length: 131,072 tokens - Vocabulary: 156,940 tokens (including 28,672 audio tokens) - License: Apache 2.0 - Voices: 8 (Nova, Aurora, Stellar, Atlas, Orion, Luna, Phoenix, Ember) To use Continue-TTS, install the required dependencies: For easier usage with audio generation, use the Continue-TTS package: Continue-TTS includes 8 professionally designed voices: | Voice | Gender | Description | |-------|--------|-------------| | nova | Female | Conversational and natural, perfect for general use | | aurora | Female | Warm and friendly, excellent for storytelling | | stellar | Female | Energetic and bright, great for upbeat content | | atlas | Male | Deep and authoritative, ideal for narration | | orion | Male | Friendly and casual, perfect for conversational content | | luna | Female | Soft and gentle, excellent for calm narration | | phoenix | Male | Dynamic and expressive, great for engaging content | | ember | Female | Warm and engaging, perfect for emotional expression | Supported emotions: - ` ` - Natural laughter - ` ` - Light laugh - ` ` - Expressive sigh - ` ` - Surprised gasp - ` ` - Cough sound - ` ` - Yawn - ` ` - Groan - ` ` - Sniffle - Audiobook Narration: Natural storytelling with emotional expression - Virtual Assistants: Conversational AI with personality - Accessibility: Text-to-speech for visually impaired users - Content Creation: Voiceovers for videos, podcasts, and presentations - Gaming: Dynamic character voices and dialogue - Education: Interactive learning materials with voice - Customer Service: Natural-sounding automated responses - Quality: State-of-the-art natural speech synthesis - Latency: ~200ms for streaming generation (GPU) - Speed: Real-time on GPU, slower on CPU - Memory: ~7GB GPU RAM (FP16), ~14GB (FP32) - Sample Rate: 24kHz (high quality audio) Continue-TTS is built on the Continue-1-OSS and combines: - Base Model: Continue-1-OSS (LLaMA-based, 3.3B parameters) - Audio Codec: SNAC multi-scale neural audio codec - Token Structure: 7 audio tokens per frame (hierarchical encoding) - Training: Fine-tuned on few hours of diverse speech data The model generates audio tokens autoregressively, which are then decoded into waveforms using the SNAC neural codec. Continue-TTS was fine-tuned on the Continue-1-OSS using: - High-quality speech datasets covering diverse accents and styles - Multi-speaker recordings for voice diversity - Emotional speech data for expressive synthesis - Conversational and narrative content Training utilized: - Continue-1-OSS as base - Custom tokenizer with 28,672 audio tokens - Multi-stage training (pretraining + fine-tuning) - Optimized for naturalness and emotion As with any TTS model, Continue-TTS has certain limitations: - Pronunciation: May struggle with unusual names, technical terms, or non-English words - Consistency: Long-form generation may have minor quality variations - Accents: Primarily trained on specific accent patterns - Compute: Requires GPU for real-time generation (CPU is slower) - Language: Currently optimized for English SVECTOR is committed to responsible AI development. Users should: - Transparency: Disclose when audio is AI-generated - Consent: Do not clone voices without explicit permission - Verification: Implement safeguards against deepfakes and misinformation - Attribution: Credit the model when used in public projects - Responsible Use: Avoid generating harmful, deceptive, or illegal content This model is released under the Apache License 2.0. See the LICENSE file for complete details. Continue-1-OSS builds upon advances in neural speech synthesis, large language models, and neural audio codecs. We thank the open-source community for their contributions to these foundational technologies.