Cognition has released SWE-2, the coding model it built by post-training Moonshot AI’s Kimi K3 with reinforcement learning.
The company behind the Devin agent reports a 50.0 percent score on FrontierCode 1.1 Main, within one point of Anthropic’s Claude Fable 5.1, at 64 percent lower cost. SWE-2 is also the first Cognition model with selectable reasoning-effort levels, and the company says all of them came out of a single training run.
The release says something about the open-weight stack. Cognition did not train a base model. It took a 2.8T-parameter open system from a Chinese lab and layered its own agentic behavior on top, an approach far cheaper than pretraining.
The model is not available for self-hosting, so customers buy it through Cognition’s own service rather than running weights on their own servers. That keeps the tuning private but ties adoption to a single vendor.
The move follows a wave of similar bets. Agent startups increasingly compete on post-training and harness design instead of model scale, a pattern that lowers the cost of entering the coding market and unsettles labs selling frontier access.
