Code World Model is Meta's 32 billion parameter language model. It simulates execution of Python functions and reproduces agentic interactions in a Bash environment. It does more than predict the next line of code. It thinks ahead, simulates, and plans the way a human would. Listed scores are 68.6% on LiveCodeBench v5, 76% on AIME24, and 65.8% on SweBench Verified with test-time scaling. Linked resources include the tech report, model weights, Hugging Face checkpoints, and inference code.
Overview
- Code World Model (CWM): Meta's 32-billion-parameter language model
- It does more than predict the next line of code:
- Simulates the execution of Python functions
- Reproduces agentic interactions in a Bash environment
- Significance: AI that doesn't just write code but also thinks ahead, simulates, and plans the way a human would
Results
- 68.6% on LiveCodeBench v5
- 76% on AIME24
- 65.8% on SweBench Verified (with test-time scaling)
Applications and Potential
- Test how a world model improves code generation
- Build AI agents that can reason and plan
- Bring coding AI closer to how humans actually think
- Open up an "open lab" for the research community
Other World Models
- A self-supervised world model trained on video, enabling zero-shot robot control and visual understanding
Resources
- Tech Report: https://ai.meta.com/research/publications/cwm
- Model Weights: https://ai.meta.com/.../models-and-libraries/cwm-downloads/
- HuggingFace: https://huggingface.co/facebook/cwm
- HuggingFace SFT: https://huggingface.co/facebook/cwm-sft
- HuggingFace Pretrain: https://huggingface.co/facebook/cwm-pretrain
- Inference Code: https://github.com/facebookresearch/cwm
See also: LLM Overview · Coding Agents