joooelw/BEHAVE-27B
joooelw/BEHAVE-27B is a 27 billion parameter model developed by joooelw, fine-tuned from Qwen3.8-27B. This model is specifically designed for multi-turn co-development of Register Transfer Level (RTL) designs and executable behavior models for verification. It excels in tasks related to hardware design and verification, offering specialized capabilities for generating and refining RTL and behavioral models.
Loading preview...
BEHAVE-27B: RTL and Behavioral Model Co-Development
BEHAVE-27B is a specialized 27 billion parameter language model, fine-tuned from Qwen3.8-27B, focusing on the multi-turn co-development of Register Transfer Level (RTL) designs and executable behavior models for verification. This particular release is the HF40 checkpoint, representing the model after 40 GRPO updates using a fixed 540-task BEHAVE-Train pool.
Key Capabilities
- RTL Design Co-development: Assists in the iterative creation and refinement of hardware designs at the RTL level.
- Executable Behavior Model Generation: Capable of generating behavioral models that can be used for verification purposes.
- Multi-turn Interaction: Optimized for conversational and iterative development workflows, allowing for continuous refinement based on user input.
- Specialized for Hardware Verification: Tailored to the unique requirements of hardware design and verification tasks.
Important Considerations
- Origin: Fine-tuned from Qwen3.8-27B, inheriting its foundational architecture.
- Limitations: Generated RTL and reference models may contain errors and require independent verification. The model is not a substitute for formal hardware verification or sign-off processes.
- Usage: Requires a model runtime compatible with Qwen3.8-27B. Reproducing multi-turn agent evaluations necessitates the BEHAVE tool environment and task settings, beyond just downloading the model weights.
For more details on the code and benchmark information, refer to the joel-wu/BEHAVE GitHub repository.