janhq

[](https://github.com/menloresearch/deep-research) [](https://opensource.org/licenses/Apache-2.0) [](https://jan.ai/) We have released a small weight update, jan-v1-2509, which refines the original v1. No architectural changes. Slightly lower performance on SimpleQA compared to jan-v1. Slightly mproved results on other chat benchmarks and overall more reliable Jan-v1 is the first release in the Jan Family, designed for agentic reasoning and problem-solving within the Jan App. Based on our Lucy model, Jan-v1 achieves improved performance through model scaling. Jan-v1 uses the Qwen3-4B-thinking model to provide enhanced reasoning capabilities and tool utilization. This architecture delivers better performance on complex agentic tasks. Question Answering (SimpleQA) For question-answering, Jan-v1 shows a significant performance gain from model scaling, achieving 91.1% accuracy. The 91.1% SimpleQA accuracy with Jan-v1 remains a highlight, though Jan-v1-2509 focuses on balancing factual QA with improved reliability across chat-based reasoning tasks. These benchmarks evaluate the model's conversational and instructional capabilities. Jan-v1 is optimized for direct integration with the Jan App. Simply select the model from the Jan App interface for immediate access to its full capabilities. - Discussions: HuggingFace Community - Jan App: Learn more about the Jan App at jan.ai () Note By default we have system prompt in chat template, this is to make sure the model having the same performance with the benchmark result. You can also use the vanilla chat template without system prompt in the file chattemplateraw.jinja.

Jan-v1-edge: Distilled for Edge, Built for Web Search [](https://github.com/menloresearch/deep-research) [](https://opensource.org/licenses/Apache-2.0) [](https://jan.ai/) 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. Despite having only 1.7B parameters, Jan-v1-edge achieves 83% accuracy—nearly matching the larger Jan-nano-128k—demonstrating its efficiency and robustness. Versus Qwen 3 1.7B Thinking, Jan-v1-edge shows a slight degradation on instruction-following and CreativeWriting, while remaining comparable or better on EQBench and recency QA. Jan-v1-edge is optimized for direct integration with the Jan App. Simply select the model from the Jan App interface for immediate access to its full capabilities. - Discussions: HuggingFace Community - Jan App: Discover more about the Jan App at jan.ai

[](https://github.com/menloresearch/deep-research) [](https://opensource.org/licenses/Apache-2.0) [](https://jan.ai/) We have released a small weight update, jan-v1-2509, which refines the original v1. No architectural changes. Slightly lower performance on SimpleQA compared to jan-v1. Slightly mproved results on other chat benchmarks and overall more reliable Jan-v1 is the first release in the Jan Family, designed for agentic reasoning and problem-solving within the Jan App. Based on our Lucy model, Jan-v1 achieves improved performance through model scaling. Jan-v1 uses the Qwen3-4B-thinking model to provide enhanced reasoning capabilities and tool utilization. This architecture delivers better performance on complex agentic tasks. Question Answering (SimpleQA) For question-answering, Jan-v1 shows a significant performance gain from model scaling, achieving 91.1% accuracy. The 91.1% SimpleQA accuracy represents a significant milestone in factual question answering for models of this scale, demonstrating the effectiveness of our scaling and fine-tuning approach. These benchmarks evaluate the model's conversational and instructional capabilities. Jan-v1 is optimized for direct integration with the Jan App. Simply select the model from the Jan App interface for immediate access to its full capabilities. - Discussions: HuggingFace Community - Jan App: Learn more about the Jan App at jan.ai () Note By default we have system prompt in chat template, this is to make sure the model having the same performance with the benchmark result. You can also use the vanilla chat template without system prompt in the file chattemplateraw.jinja.

Jan-v1-edge: Distilled for Edge, Built for Web Search [](https://github.com/menloresearch/deep-research) [](https://opensource.org/licenses/Apache-2.0) [](https://jan.ai/) 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. Despite having only 1.7B parameters, Jan-v1-edge achieves 83% accuracy—nearly matching the larger Jan-nano-128k—demonstrating its efficiency and robustness. Versus Qwen 3 1.7B Thinking, Jan-v1-edge shows a slight degradation on instruction-following and CreativeWriting, while remaining comparable or better on EQBench and recency QA. Jan-v1-edge is optimized for direct integration with the Jan App. Simply select the model from the Jan App interface for immediate access to its full capabilities. - Discussions: HuggingFace Community - Jan App: Discover more about the Jan App at jan.ai

- Developed by: janhq - License: apache-2.0 - Finetuned from model : unsloth/Qwen3-1.7B This qwen3 model was trained 2x faster with Unsloth and Huggingface's TRL library.