Jan-v1: Advanced Agentic Language Model

Jan-v1 is the inaugural release in the Jan Family of models, developed by volosati and specifically engineered for agentic reasoning and problem-solving. This 4 billion parameter model is built upon the existing Lucy model and integrates the Qwen3-4B-thinking architecture to deliver improved performance on complex agentic tasks.

Key Capabilities & Performance

• Enhanced Agentic Reasoning: Designed to excel in scenarios requiring advanced reasoning and tool utilization.
• Strong Question Answering: Achieves 91.1% accuracy on SimpleQA, demonstrating significant gains in factual question answering through model scaling and fine-tuning.
• Conversational & Instructional: Evaluated with chat benchmarks to ensure robust conversational and instructional capabilities.

Integration & Deployment

Jan-v1 is primarily optimized for seamless integration with the Jan App, allowing users to access its full capabilities directly. For developers, it supports local deployment via vLLM and llama.cpp, with recommended parameters for optimal performance. The model includes a default system prompt in its chat template to maintain benchmark performance, with an option for a vanilla chat template.