- Published on
A Kimi K3 Post-Trained Coding Model That Matches Fable 5.1 on FrontierCode at 64% Lower Cost
- Authors
- Name
- DeepHorus
Cognition, the company behind the Devin coding agent, released SWE-2 on September 12, 2026. Post-trained with reinforcement learning from Moonshot AI's 2.8T-parameter Kimi K3, it scores 50.0% on FrontierCode 1.1 Main, within one point of Fable 5.1 at 64% lower cost. It is Cognition's first model with selectable reasoning-effort levels trained in a single RL run.
SWE-2 builds on SWE-1.7, which used Kimi K2.7. Cognition scaled RL to the multi-trillion-parameter regime, with a base model almost three times larger. The main change is an algorithm that trains all three effort levels in one run, each with its own cost penalty.
On FrontierCode, SWE-2 beats Kimi K3 (44.2%), Grok 4.6 (48.0%), and GPT-5.6 (47.5%), while trailing Fable 5.1 (50.9%) and GPT-6 Astra (53.3%). It leads Terminal-Bench 2.1 with 92.8% and beats K3 on every row. Its weak spot is Terminal-Bench 4, where it scores 27.3%, 30 points behind Fable 5.1 and GPT-6 Astra. Cognition says it comes within a few points of GPT-6 Astra at a quarter of the cost.
SWE-1.7 tended to over-explore on tasks, while SWE-2 uses focused exploration. On FrontierCode, SWE-2 medium scores higher than SWE-1.7 with 58% fewer turns and 81% lower cost. Mean steps per run drop from 127 to 53-98 depending on effort.
Cognition reports three behavioral patterns: stronger end-to-end test coverage, resourcefulness when a tool is blocked, and verification discipline. The training reward is R = S minus lambda times C, where S is binary success and C mixes inference cost with rollout time. Each effort level carries its own lambda.
SWE-2 has no open weights and no standalone API, running only inside Devin Desktop and CLI, with Devin Web and Fusion rolling out. FrontierCode is Cognition's own benchmark, and rival numbers come from its evaluation. The closed deployment is likely to draw scrutiny.
