Cognition Releases SWE-2: A Kimi K3 Post-Trained Model

Cognition has officially introduced SWE-2, an advanced artificial intelligence coding model post-trained directly from the Kimi K3 base model. Building upon the infrastructure and recipe of its predecessor SWE-1.7, which was derived from Kimi K2.7, Cognition scaled its reinforcement learning (RL) into the multi-trillion-parameter regime utilizing a base model with nearly three times the parameter count. The company noted that its RL framework continues to discover substantial headroom on top of K3, driving benchmark gains by 5 to 6 points across multiple evaluations.

Regarding deployment, SWE-2 is not accessible on independent infrastructure as it features no open weights and no standalone API. Instead, it operates exclusively inside the Devin ecosystem, currently available on Desktop and CLI, with Devin Web and Fusion rolling out soon.

Stock photo for illustration only, not from the actual event

The primary architectural shift involves an RL algorithm capable of training all three effort levels in a single run, attaching an individual cost penalty to each tier so the cost-and-performance frontier shifts simultaneously. Furthermore, SWE-2 addresses the tendency of SWE-1.7 to over-explore simple assignments through a mechanism termed focused exploration. On FrontierCode 1.1 Main, the medium variant achieves higher scores while consuming 58% fewer turns and reducing costs by 81%. The mean steps per run drop significantly, and the medium model executes its first genuine edit after a median of 18 steps compared to 48 for SWE-1.7.

The Cognition team also outlined three prominent behavioral patterns: stronger end-to-end test coverage, enhanced resourcefulness during blocked tool scenarios, and strict verification discipline. When challenged, the model re-derives its conclusions rather than simply re-asserting them. In terms of data pipeline improvements, Cognition tripled its RL environments, incorporated instruction-following overlays, and established a flywheel that leverages prior SWE-2 checkpoints to correct false positives and negatives within verifiers.