Code World Model
Summary
Code World Model (CWM) is Meta FAIR’s 32B open-weight research LLM for code generation and reasoning with computational-environment world modeling. It is trained on ordinary code/text data plus explicit action-observation trajectories from Python execution traces and agentic Docker environments.
Official Artifacts
- Technical report: https://arxiv.org/abs/2510.02387
- Meta research page: https://ai.meta.com/research/publications/cwm-an-open-weights-llm-for-research-on-code-generation-with-world-models/
- Code repository: https://github.com/facebookresearch/cwm
- Hugging Face model card: https://huggingface.co/facebook/cwm
- Model downloads: https://ai.meta.com/resources/models-and-libraries/cwm-downloads/
- Preparedness report: https://ai.meta.com/research/publications/code-world-model-preparedness-report/
Role In The Wiki
CWM is the current code-domain anchor for the idea that an LLM can be trained on explicit action-observation trajectories rather than only static tokens. Its strongest transfer value is the interface pattern:
code / repo context + current computational state + action
-> next computational state or environment responseFor the Foundation TSFM agenda, CWM is adjacent but important. It shows a serious digital-world trajectory recipe, but the environment is executable code and Docker sandboxes, not numeric telemetry, graph time series, incident streams, or typed production interventions.
Boundary
CWM should not be treated as a telemetry-native action-conditioned world model. It may be a strong LLM-agent baseline or a code/repository reasoning layer for digital operations, but it does not replace a model trained on operational observations, actions, rewards, and outcomes.
The most useful analogy is:
CWM: code + program/tool action -> program/environment state
SRE world model: telemetry + graph + operator action -> future telemetry/rewardRelation To Foundation TSFM Agenda
Use the source-level agenda mapping in cwm-2025 rather than duplicating verdict rows here.
At the entity level, CWM is the model/object card. The source page carries the slot-level verdicts, evidence, limitations, and open questions.