CognitionSWE
SWE-2 Flash
SWE-2 Flash has 276 billion total and 12 billion active parameters. Cognition adds mid-training and reinforcement learning to Inkling-Small.
What the release provides
The mid-training data includes code, scientific articles and other text. Three stages place the scarcest high-quality data later. Public weights are not released; Cognition describes a future controlled sharing program for partners and researchers.Cognition: SWE-2 Flash
How to use it well
Practical guidance. Use it for a complete, bounded engineering task and inspect the result. Product access does not provide a checkpoint that can be hosted on your own systems.
Specifications & access
| Detail | Value and conditions |
|---|---|
| Access | Provider product |
| Current access | Devin Desktop and CLI; not Devin cloud. Medium is available on free plans. High and Max require a paid plan.Devin: Current model catalog |
Related models
A coding-agent model post-trained from Moonshot’s Kimi K3.
Launch date not confirmedProvider product