BEHAVE-27B

BEHAVE-27B is a model for multi-turn co-development of RTL designs and executable behavior models for verification.

This release is the HF40 checkpoint, obtained after 40 GRPO updates using the fixed 540-task BEHAVE-Train pool, starting from Qwen3.8-27B. It is the fixed-dataset RL model, not the separate self-improvement model.

Code and benchmark information: joel-wu/BEHAVE.

Checkpoint and files

• The hf40 tag identifies the paper's 40-update checkpoint. The internal export is numbered hf_39 because update indices start at zero.
• The release contains full model weights in 178 Safetensors shards, the weight index, model configuration, tokenizer, chat template, processor configuration, and the Apache 2.0 license.
• Optimizer states, private training logs, and other checkpoints are not included.

Download

from huggingface_hub import snapshot_download

model_dir = snapshot_download(
    repo_id="joooelw/BEHAVE-27B",
    revision="hf40",
)

Use a model runtime compatible with Qwen3.8-27B and the included configuration and chat template. Reproducing the multi-turn agent evaluation also requires the BEHAVE tool environment and task settings; downloading weights alone does not reproduce those experiments.

Limitations

Generated RTL and reference models can contain errors and require independent verification before use. The model is not a replacement for hardware verification or signoff.

License

Apache 2.0. See LICENSE.