Meta FAIR releases Code World Model for code generation and reasoning

New from Meta FAIR: Code World Model (CWM), a 32B-parameter research model designed to explore how world models can transform code generation and reasoning about code. We believe in advancing research in world modeling and are sharing CWM under a research license to help empower the community to build upon our work. ➡️ Read the technical report: https://lnkd.in/gJwwqiZB ➡️ Download the open weights: https://lnkd.in/gT9UvANm ➡️ Download the code: https://lnkd.in/g7RXZbwC

Code World Model (#CWM): writing code by imagining what will happen when it runs and choosing steps that create the desired result.

🫡 Great to see FAIR exploring similar directions! We're also exploring how code llms can directly serve as execution environments. Imo there are two key benefits for such models: (i) learning from tool outputs as a way for llms to learn from experience, and (ii) accelerate RL training with simulated experience instead of slow real-world interactions

🚀 Thrilled to announce the release of Meta Code World Model (CWM) — a 32B-parameter dense LLM designed to push the boundaries of code generation, reasoning, and world modeling. When humans plan, we simulate possible outcomes in our heads. Similarly, reasoning about code often means mentally executing parts of it. Current LLMs struggle here — and that’s where CWM comes in. Trained on extensive coding data plus bespoke Python + Bash “world modeling” data, CWM can simulate function execution and agentic interactions, opening up new research directions for code world models. Key results: 📊 LiveCodeBench v5: 68.6% 📊 AIME24: 76% 📊 SweBench Verified: 65.8% (with test-time scaling) We’re releasing CWM openly for research to help the community build on this work: 📊 Tech Report https://lnkd.in/eV7YirjC ⚖️ Models weights https://lnkd.in/e2CTzsxr 🤗 On Huggingface https://lnkd.in/e_S4R-P4 https://lnkd.in/eWDgyCEs https://lnkd.in/ex47F4sd 🧑💻 Inference Code https://lnkd.in/eVHeW8VV This has been an incredible team effort across AI Research at Meta — with significant contributions from researchers, engineers, PhD students, and veterans alike, led by Gabriel Synnaeve, and including Taco Cohen, Yossef Adi, Jade Copet, Michel Meyer, and many others! Huge thanks to everyone involved in making this happen! We’re excited to see what the community will build with CWM, and how world models might reshape the future of code generation. #MetaAI #CodeWorldModel #LLM #WorldModel #AIResearch #CodeGeneration

Meta just released Code World Model (CWM): a 32B parameter LLM trained not just on code but also on Python + Bash world modeling data. It can simulate function execution and reason in agentic environments. This opens up new research on how world models can improve planning and reasoning for code generation. As I’ve been following these developments, one thing is clear: world models aren’t just for code. In embodied AI, they’re becoming central to how agents learn, plan, and act. I pulled my thoughts together here: https://lnkd.in/esizeHF3 Let me know what you think!

From Guessing to Simulating: A Step Toward AGI Imagine learning a video game. With an LLM, you ask a friend every time: “What happens if I press this button?” — slow and repetitive. With a Code World Model, the friend writes you a mini-simulator you can play with endlessly — fast, clear, and reusable. This shift means AI won’t just guess outcomes, it will build world models we can inspect, trust, and use for smarter decisions. It’s not just about robotics or planning — it’s a glimpse at how AI could start reasoning like us, a step toward true AGI.

Meta FAIR: Code World Model (CWM), a 32B-parameter research model designed to explore how world models can transform code generation and reasoning about code.

Meta FAIR’s new Code World Model (32B) is an ambitious step toward bringing world models into code reasoning. The idea of moving beyond token prediction to actually model program states and execution paths is powerful — but the big question is whether this approach will scale from research into practical developer tools. Excited to see where the experiments lead, but cautious on how soon this translates into day-to-day coding impact.

Meta FAIR released Code World Model (CWM), a 32-billion-parameter research model. It is designed to investigate the application of world models for code generation and code reasoning. The model's weights and code are being shared under a research license to advance community research.

Meta's announcement of the Code World Model (CWM) represents an exciting advancement in the field of AI, particularly in relation to code generation and reasoning. With 32 billion parameters, this research model is positioned to explore the potential of world models in transforming how we understand and generate code.