SWE-2 Flash is Cognition’s new coding model, a faster and cheaper sibling of its larger SWE-2.

It’s 276B parameters in total, with 12B active, built on Inkling-Small from Thinking Machines. Cognition says it generates about 200 tokens per second, 3x the speed of SWE-2, at one-fifth the price.

Gains over the base model:

Terminal-Bench 4.0: 0.9% to 20.0%
FrontierCode 2.0: 3.2% to 24.0%
ExploitBench (security exploit tasks): 11.3% to 34.8%

For context, Cognition’s own table shows other models scoring higher on some of these. For example, GLM 5.3 Flash and DeepSeek V4.1 Flash 1017 are above 41% on Terminal-Bench 4.0, and SWE-2 itself scores higher on most rows.

On training, Cognition added a mid-training stage (a few trillion tokens of code, articles and blogs) before reinforcement learning, for the first time in its recipe. It says most of the gains came from higher-quality and more diverse RL tasks.

You can try it in Devin Desktop and the Devin CLI, free through November 30. The weights aren’t public yet. Cognition says it’s building a program to share them with trusted industry partners and research labs, and will say more in a few weeks.

Good fit if you want a fast, low-cost coding model inside Devin’s tools and don’t need the largest model for every task.

SWE-2 Flash is Cognition’s new coding model, a faster and cheaper sibling of its larger SWE-2.

Gains over the base model:

Terminal-Bench 4.0: 0.9% to 20.0%
FrontierCode 2.0: 3.2% to 24.0%
ExploitBench (security exploit tasks): 11.3% to 34.8%

Good fit if you want a fast, low-cost coding model inside Devin’s tools and don’t need the largest model for every task.

Cognition SWE-2 Flash

Cognition’s fast coding model, 3x the speed of SWE-2

Cognition’s fast coding model, 3x the speed of SWE-2