Cognition launches a cheaper, faster coding model — but keeps the weights private
SWE-2 Flash runs inside Devin at one-fifth the price and three times the speed of SWE-2, while cyber-risk concerns limit access to its weights.
Cognition has released SWE-2 Flash, a coding model designed to make everyday agent work faster and cheaper. The company says it is available in Devin Desktop and CLI at one-fifth the price and three times the generation speed of its larger SWE-2 model.
SWE-2 Flash is built on Thinking Machines Lab’s Inkling-Small, a mixture-of-experts model with 276 billion total parameters and 12 billion active parameters. Cognition added a mid-training stage using several trillion tokens of code, scientific articles, blogs and other material, followed by large-scale reinforcement learning.
The company reports that the training lifted Terminal-Bench 4.0 from 0.9% to 20.0% and FrontierCode 2.0 from 3.2% to 24.0% versus the base model. Those are vendor-run results, and the comparison is mixed: SWE-2 Flash leads the American open-weight models cited by Cognition, but GLM 5.3 Flash and DeepSeek V4.1 Flash score much higher on Terminal-Bench and DeepSWE.
The unusual part is access. Cognition says cybersecurity training raised ExploitBench performance from 11.3% to 34.8%, improving defensive usefulness while also increasing misuse risk. It is therefore not releasing the weights publicly yet. Instead, the company plans a controlled-access program for trusted industry partners and research labs.
For Devin users, the practical change is lower-cost coding automation for code review, feature work, documentation and routine engineering tasks. Eligible paid plans can use SWE-2 Flash free for the rest of the month.
Cognition
