Cognition released SWE-2, its newest coding model. The model is available today in Devin Desktop and the Devin command-line tool. Cognition is also rolling it out on Devin Web and Fusion. SWE-2 is built on Kimi K3, an existing open-weight coding model. Cognition trained all of its effort levels in one reinforcement learning run. It says a cost penalty in training pushed both price and capability together. SWE-2 takes far fewer steps per task than SWE-1.7 and starts editing code sooner. Cognition reports better test writing, more persistence when a path is blocked, and less agreement with wrong user claims. It also reran two trustworthiness checks covering political censorship and insecure code. No customer or language framing changed how vulnerable the generated code was.
What changed
Cognition's previous model, SWE-1.7, scored 42.0% on the same test and over-explored simple tasks.
What it unlocks
Running Devin's coding agent at a chosen effort level, with lower cost per task.
- 50.0% on FrontierCode 1.1 Main
- 64% cheaper than Fable 5.1
- 81% lower cost than SWE-1.7
- 2.8T-parameter base model
What you need to act on it
- Devin Desktop or Devin CLI
Sources