Cognition releases SWE-2 Flash, a coding model built on Thinking Machines' Inkling-Small
The 276B-parameter model with 12B active is offered in Devin at one-fifth the price and 3x the speed of SWE-2; Cognition's own evals lift Terminal-Bench 4.0 from 0.9% for the base model to 20.0%, and the weights are not public.
Why it matters
SWE-2 Flash is one of the first commercial models built on Thinking Machines' Inkling-Small, and Cognition says adding a mid-training stage before reinforcement learning drove most of the gain over that base. It is pitched as a cheap, fast worker for coding agents rather than a frontier model: the larger SWE-2 still scores higher on most of Cognition's own benchmarks. Cognition says it will share weights only with vetted partners because of cyber risk. Watch for independent benchmark runs and whether other labs build on Inkling.
Key facts
- 276B total parameters, 12B active
- Base model: Inkling-Small from Thinking Machines
- Terminal-Bench 4.0: 20.0% vs 0.9% for the base model
- FrontierCode 2.0: 24.0% vs 3.2% for the base model
- One-fifth the price and 3x the speed of SWE-2
- Available in Devin Desktop and CLI; weights not public