HeadsUpAI

Cognition Launches SWE-check to Match Frontier Bug Detection at 10x Speed

· Updated

Cognition, an applied AI lab building end-to-end software agents, released SWE-check with Applied Compute. This specialized model identifies bugs in code diffs and generates structured fixes. It was trained using reinforcement learning (learning through trial and error) to match the accuracy of frontier generalists like Opus 4.6 while running ten times faster.

General-purpose models often suffer from high latency that breaks a developer's focus. SWE-check avoids this by using a two-phase training approach: first maximizing bug-finding capability, then applying a latency penalty based on real user data to compress the model's reasoning without sacrificing its specialized skill.

You can access a preview of SWE-check today in Windsurf Next using the cmd+U shortcut, with a mainstream release for Windsurf coming soon. The model provides structured bug descriptions and fixes directly in the IDE, enabling a near-real-time experience that keeps you in flow.

View the full update on cognition.ai
Cognition
Cognition
@cognition
X

Today we're releasing SWE-check, a specialized bug detection model we RL-trained with @appliedcompute that matches frontier performance on internal in-distribution evals and makes meaningful progress on out-of-distribution evals, all while running 10x faster. https://t.co/PojZxeiCaa

30retweets266likes
View on X

Share this update