Code World Model (CWM) 32B

2025-09-24 · Publisher : Meta · Domain : Artificial Intelligence

Meta FAIR releases Code World Model (CWM), a dense 32-billion-parameter open-weights model for research, trained to simulate code execution and to solve software engineering tasks.

Meta FAIR releases Code World Model (CWM), a dense 32-billion-parameter open-weights model for research, trained to simulate code execution and to solve software engineering tasks. ### Context After Llama 4, Meta had no new general-purpose language model. CWM is a research model from FAIR, released under a non-commercial license, in a market where Qwen3 and Devstral dominated open models of about 30 B for code. It precedes the closed turn of Meta Superintelligence Labs. ### What the model brings - Mid-training on Python interpreter execution traces and agent interactions in Docker containers (about 3 million trajectories according to MarkTechPost). - Post-training through multitask reinforcement learning (verifiable code, mathematics, multi-turn software engineering). - Release of three sets of weights (pre-training, SFT, post-RL) for research. - Goal: simulate the state of variables step by step to better plan and fix code. ### Architecture and specifications Dense 32 B transformer, decoder-only, 64 blocks, GQA, alternating local (8,192 tokens) and global (131,072 tokens) attention at a 3:1 ratio, 128 K vocabulary. Context of 131,072 tokens. Language: English only according to the license. Data: about 8 trillion tokens of code-heavy pre-training, then 5 trillion tokens of mid-training (MarkTechPost). License: FAIR Non-Commercial Research License (commercial use prohibited). Weights on Hugging Face (facebook/cwm), technical report arXiv 2510.02387. Runs on an 80 GB GPU with quantization according to the model card. ### Cost and access | Item | Value | |---|---| | Official API | None (weights only) | | Hardware | 1 80 GB GPU with quantization (model card) | | Commercial use | Prohibited by the license | ### Benchmarks | Benchmark | CWM 32B | Qwen3-32B | Magistral-small | |---|---|---|---| | SWE-bench Verified | 53.9 (65.8 with test-time scaling) | not disclosed | not disclosed | | LiveCodeBench v5 | 68.6 | 65.7 | 70.0 | | Math-500 | 96.6 | 97.2 | not disclosed | | AIME 2024 | 76.0 | 81.4 | 86.1 | | AIME 2025 | 68.2 | 72.9 | 77.3 | Source: Hugging Face model card facebook/cwm, scores published by the vendor; comparison only with open models. No comparison table with closed models is published. ### Versus closed models Meta writes that CWM is competitive with open models of comparable size and even with larger or closed models on SWE-bench Verified: this is a vendor claim (cited by MarkTechPost), with no quantified table against closed models in the sources consulted. The 65.8% score is only reached with test-time scaling, versus 53.9% with direct generation. On mathematics, CWM stays behind Qwen3-32B (AIME 2024: 76.0 versus 81.4). The non-commercial license rules out any production use, which restricts it to research. It therefore does not change a company's choice between open and closed, but documents a technical direction from FAIR. ### Key takeaways CWM is Meta's last notable open model before Muse Spark: a strong research model for code, but not commercially usable.

Tags : LLM, Open Source, Meta, Code, FAIR Non-Commercial Research License, Recherche, Agents