CWM: An Open-Weights LLM for Research on Code Generation with World Models (ai.meta.com)

The team released Code World Model (CWM), a 32B parameter, open‑weights decoder‑only LLM designed to advance research in code generation by training on dynamic “world” traces rather than static source code alone. CWM is mid‑trained on large corpora of observation–action trajectories collected from a Python interpreter and agentic Docker environments, then refined with extensive multi‑task reasoning RL across verifiable coding, math, and multi‑turn software‑engineering tasks. It supports very long contexts (up to 131K tokens) and demonstrates strong standalone performance: pass@1 of 65.8% on SWE‑bench Verified (with test‑time scaling), 68.6% on LiveCodeBench, 96.6% on Math‑500, and 76.0% on AIME 2024. Why this matters: CWM provides a practical testbed showing how world models—models that learn to simulate an execution environment—can improve agentic coding, stepwise execution simulation, planning, and reasoning about code behavior. Training on observation/action trajectories supplies explicit causal signals about execution and interaction, RL on verifiable tasks aligns behavior to correctness, and the long context enables multi‑turn debugging and complex trace reasoning. Crucially, the authors release intermediate checkpoints (post mid‑train, SFT, and RL) to spur reproducible research into agentic code generation, environment simulation, and the limits of world‑modeling techniques.