Highlights 🎉
Jan v0.6.7 brings full support for OpenAI’s groundbreaking open-weight models - gpt-oss-120b and gpt-oss-20b - along with enhanced MCP documentation and critical bug fixes for reasoning models.
🚀 OpenAI gpt-oss Models Now Supported
Jan now fully supports OpenAI’s first open-weight language models since GPT-2:
gpt-oss-120b:
- 117B total parameters, 5.1B active per token
- Runs efficiently on a single 80GB GPU
- Near-parity with OpenAI o4-mini on reasoning benchmarks
- Exceptional tool use and function calling capabilities
gpt-oss-20b:
- 21B total parameters, 3.6B active per token
- Runs on edge devices with just 16GB memory
- Similar performance to OpenAI o3-mini
- Perfect for local inference and rapid iteration
🎮 GPU Layer Configuration
Due to the models’ size, you may need to adjust GPU layers based on your hardware:
Start with default settings and reduce layers if you encounter out-of-memory errors. Each system requires different configurations based on available VRAM.
📚 New Jupyter MCP Tutorial
We’ve added comprehensive documentation for the Jupyter MCP integration:
- Real-time notebook interaction and code execution
- Step-by-step setup with Python environment management
- Example workflows for data analysis and visualization
- Security best practices for code execution
- Performance optimization tips
The tutorial demonstrates how to turn Jan into a capable data science partner that can execute analysis, create visualizations, and iterate based on actual results.
🔧 Bug Fixes
Critical fixes for reasoning model support:
- Fixed reasoning text inclusion: Reasoning text is no longer incorrectly included in chat completion requests
- Fixed thinking block display: gpt-oss thinking blocks now render properly in the UI
- Fixed React state loop: Resolved infinite re-render issue with useMediaQuery hook
Using gpt-oss Models
Download from Hub
All gpt-oss GGUF variants are available in the Jan Hub. Simply search for “gpt-oss” and choose the quantization that fits your hardware:
Model Capabilities
Both models excel at:
- Reasoning tasks: Competition coding, mathematics, and problem solving
- Tool use: Web search, code execution, and function calling
- CoT reasoning: Full chain-of-thought visibility for monitoring
- Structured outputs: JSON schema enforcement and grammar constraints
Performance Tips
- Memory requirements: gpt-oss-120b needs ~80GB, gpt-oss-20b needs ~16GB
- GPU layers: Adjust based on your VRAM (start high, reduce if needed)
- Context size: Both models support up to 128k tokens
- Quantization: Choose lower quantization for smaller memory footprint
Coming Next
We’re continuing to optimize performance for large models, expand MCP integrations, and improve the overall experience for running cutting-edge open models locally.
Update your Jan or download the latest.
For the complete list of changes, see the GitHub release notes.