RL hasn't peaked yet: SWE-1.7

Cognition's SWE-1.7 code model uses reinforcement learning on Moonshot AI's Kimi K2.7 to score 42.3% on FrontierCode Main, now live in Devin via Cerebras.

SWE-1.7, Cognition's (the developer of the Devin agent) new code model, arrives with a thesis that goes beyond a simple launch: reinforcement learning (RL) is far from having reached its ceiling.

Starting from an unusual point, training begins with an open model, Kimi K2.7 from Moonshot AI, which has already undergone extensive post-training. On this foundation, Cognition claims to have achieved significant gains through its own RL, enough to contradict the idea of a "post-training ceiling" and suggest that the method extends much further than previously thought.

The infrastructure does the rest. Training was distributed across four data centers and three continents: only the trainer requires a high-bandwidth cluster, while the inference engines that produce the rollouts run anywhere and synchronize via compressed weight deltas deposited in object storage. Cognition also documents its safeguards against two pitfalls that typically stall long RL sessions: entropy collapse and numerical drift between training and inference.

Another component of the system is self-compaction: as the context limit approaches, the agent summarizes its working state and resumes from that summary, extending some rollouts up to six hours. The developer observes in return a model that explores the codebase more thoroughly before acting, at the cost of a wider perimeter of modified files.

On its own FrontierCode Main benchmark, Cognition claims a score of 42.3% for an average cost of approximately two dollars per task. SWE-1.7 is available today in Devin (web, desktop, CLI), served at 1,000 tokens per second via Cerebras.