license: apache-2.0
language:
- en
base_model:
- Qwen/Qwen3-4B-Thinking-2507
pipeline_tag: text-generation
Jan-v1: The Inaugural Model of the Jan Family – Redefining Agentic Reasoning
Authors: Alan Dao, Bach Vu Dinh, Alex Nguyen, Norapat Buppodom # Assuming same authors, adjust if needed.
Overview
Introducing Jan-v1, the foundational model in the Jan Family – a new lineage of highly capable language models developed to power the next generation of intelligent agents within the Jan App ecosystem. Building on the innovative agentic capabilities of our earlier Lucy model, Jan-v1 represents a significant leap forward through strategic model scaling.
By leveraging a larger Qwen3-4B base, Jan-v1 demonstrates profoundly enhanced 'thinking' and reasoning capabilities. This architectural evolution is designed to deliver superior performance on complex agentic tasks, setting a new benchmark for accessible, high-performance AI.
What Jan-v1 Excels At
- 🧠 Enhanced Agentic Reasoning: With its larger parameter count, Jan-v1 excels at deeper reasoning, complex problem-solving, and sophisticated multi-step agentic planning.
- 🎯 Superior Question Answering: Achieves an impressive 91.2% accuracy on SimpleQA, significantly advancing performance for factoid question answering.
- 🔍 Advanced Agentic Web Search: Inherits and refines Lucy's strong capabilities for agentic web search and lightweight browsing via MCP-enabled tools.
- 📱 Optimized for Jan App: Specifically engineered to provide unique and highly optimized support for the Jan App, ensuring seamless integration and superior user experience.
Evaluation
Jan-v1's strategic scaling has resulted in a notable performance uplift, particularly evident in its "thinking" and reasoning prowess. Following the established MCP benchmark methodology, Jan-v1 sets a new standard for models in its class.
| Model | SimpleQA Accuracy |
|---|---|
| Jan-v1 (Qwen3-4B) | 91.2% |
| Lucy (Qwen3-1.7B) | [Lucy's Score] |
| DeepSeek-v3 (Comparison from Lucy) | [DeepSeek's Score] |
The 91.2% accuracy on SimpleQA underscores Jan-v1's advanced ability to precisely retrieve and synthesize information, showcasing the effectiveness of our model scaling approach for agentic intelligence.
🖥️ How to Run Locally
Jan-v1 is designed for flexible deployment, compatible with various inference engines including vLLM, llama.cpp, and local applications like Jan and LMStudio. Its integration with search APIs and web browsing tools is facilitated through the MCP (Mobile-Cloud Protocol).
Deployment
Deploy using VLLM:
vllm serve Menlo/Jan-v1 \ # Update with your HF model ID (e.g., Menlo/Jan-v1)
--host 0.0.0.0 \
--port 1234 \
--enable-auto-tool-choice \
--tool-call-parser hermes
Or llama-server from llama.cpp:
llama-server ...
Recommended Sampling Parameters
Temperature: 0.7
Top-p: 0.9
Top-k: 20
Min-p: 0.0
🤝 Community & Support
- Discussions: HuggingFace Community
- Jan App: Learn more about the Jan App at jan.ai
📄 Citation
Updated Soon