Cognition releases SWE-2 Flash, a faster, cheaper coding model built on Thinking Machines' Inkling-Small

Cognition introduced SWE-2 Flash, a coding model based on Thinking Machines' Inkling-Small (276B total and 12B active parameters). Cognition says it costs 5x less and runs 3x faster than SWE-2, and that on leading evals it outperforms all American open-weight models, including ones three times larger. These are Cognition's own claims.

Cognition says it added a mid-training stage before reinforcement learning for the first time, on several trillion tokens of code, scientific articles and blogs, then trained with large-scale RL. It says most of the gains over SWE-2 came from higher-quality, more diverse RL tasks, whose instructions, environments and verifiers its researchers audited.

SWE-2 Flash is free until the end of November for Pro, Max and Teams users in Devin Desktop and Devin CLI. Cognition says it is building a program to share the weights securely with partners and researchers, to balance transparency and cyber risks, and will give details in the coming weeks.