Code World Model

Code World Model (CWM) is a 32-billion-parameter dense language model built by the Meta FAIR CodeGen team for research on code generation and code reasoning. To better represent how code and commands change program and system state, it was mid-trained on Python execution traces and agent interaction data from containerized environments. It was then post-trained with multi-task reinforcement learning in verifiable coding, math, and multi-turn software engineering environments. It is released under a non-commercial research license, access requires agreeing to provide contact information, and when quantized it can run on a single 80GB GPU.

• 32-billion-parameter dense decoder model
• Mid-trained on Python execution traces and agent interaction data
• Multi-task reinforcement learning for coding, math, and software engineering
• Post-trained, SFT, and pretrained checkpoints released
• Runs on a single 80GB GPU when quantized
• vLLM and Fastgen usage examples

Open source (free)

CWM은 비상업 연구용 라이선스(fair-noncommercial-research-license)로만 공개되며, 상업 제품·서비스 사용은 금지되고 연락처 제공 동의 후 접근 가능

공식 페이지 확인 2026-10-14(산출 규칙 2026-10-15)

• One of 60 AI tools in the Foundation Models field.
• Connected to 17 other tools — #1 in its field.