Home » Technology » facebook/cwm · Hugging Face

facebook/cwm · Hugging Face

by Rachel Kim – Technology Editor

Facebook AI’s Code World Model⁣ Achieves State-of-the-Art ‍Results on Coding Benchmarks

MENLO PARK, CA ​- ⁢September 25, 2024, 01:21:04 PDT – Facebook AI has released ⁤Code World Model (CWM), a new large language model demonstrating‍ competitive performance​ on several established coding​ benchmarks. The model,detailed‍ in a recently published tech report,showcases strong capabilities in code generation and problem-solving,positioning it as a significant advancement in ‍the field of artificial intelligence for software progress.

CWM⁤ addresses the growing demand for AI tools capable ⁣of assisting developers⁣ with complex coding ‌tasks. Its release ‍comes as the software industry⁣ increasingly seeks automation and efficiency gains through ‌AI-powered solutions. The model’s performance suggests potential applications ranging from automated code completion and ⁣bug ‌fixing​ to the generation of entire software components, impacting developers, software companies, ‍and ultimately, the pace of technological ⁢innovation. CWM’s architecture leverages reinforcement learning techniques ‍to improve ⁣its coding ​proficiency.

Evaluations reveal CWM achieves ‌a‌ score of 68.6‍ on the LCBv5 ⁤benchmark, 63.5 on LCBv6, ⁤and 96.6 on Math-500. It also ⁣demonstrates proficiency⁤ on AIME24 (76.0) and AIME25 (68.2). When compared⁢ to other models, CWM’s 96.6​ score on Math-500 is comparable to ⁢Qwen3-32B’s 97.2, while surpassing⁤ Magistral-small-2509-24B and both low ‌and medium configurations of gpt-oss-20B.

Further testing on ‍the SweBench Verified benchmark shows CWM achieving a score of 53.9, increasing to 65.8 when combined‌ with text-to-speech (tts) functionality. This performance is on par with Devstral-1.1-2507-24B (53.6) and ⁣Qwen3-Coder-32B⁤ (51.6),‍ and exceeds⁤ the range of 37.4 to⁢ 60.7 achieved by different configurations of gpt-oss-20B.Notably, CWM was⁢ evaluated on the full 500 problems of SweBench Verified, while GPT-5 and GPT-oss utilized a ⁣custom subset of 477 problems. ⁢

The Code World Model tech report is⁣ available​ on the AI ‍at Facebook‍ research publications website.

You may also like

Leave a Comment

This site uses Akismet to reduce spam. Learn how your comment data is processed.