Overview

• Meta published three checkpoints of Code World Model to invite community research rather than to ship a production system.
• CWM was trained on large programming corpora plus specialized Python and Bash world‑modeling data to enable simulated execution and agent interaction.
• Initial results reported publicly include 65.8% on SWE‑bench Verified and 68.6% on LiveCodeBench, with 96.6% on Math‑500 and 76.0% on AIME 2024.
• Performance places the 32B model among leading open‑source code models but behind some top competitors such as Qwen3‑Coder and Kimi‑K2‑Instruct.
• Yann LeCun amplified the announcement on social media as community figures, including Alexandr Wang, shared benchmark snapshots.