Cognition launches SWE-2 Flash, a faster, cheaper coding model for Devin Desktop and CLI

Devin Central illustration

Flash runs three times faster than SWE-2 at one‑fifth the cost.

Cognition released SWE-2 Flash on Nov. The company makes the model available at no charge in Devin Desktop and Devin CLI through Nov. 30, 2026.

According to a post on orcarouter.ai, SWE-2 Flash is a mixture‑of‑experts model built on the Apache‑2.0‑licensed Inkling‑Small base from Thinking Machines Lab. It has 276 billion total parameters but only 12 billion active at inference, a contrast to SWE-2’s 2.8‑trillion‑parameter architecture. Cognition’s own size chart lists Flash at “1x” and SWE-2 at “10.1x” in active parameters, an eightfold difference.

Startupfortune.com reported that Cognition says Flash runs three times faster than SWE-2 and costs one‑fifth as much. The speed claim is based on a simulated replay of Terminal‑Bench 4.0’s gsea‑proteomics task at 200 tokens per second for Flash versus 60 tokens per second for SWE-2. Unite.ai noted that Cognition’s internal benchmarks show Flash edging SWE-2 on FrontierCode 2.0, while SWE-2 scores higher on Terminal‑Bench 4.0 and DeepSWE 1.1.

Cognition also claims Flash “outperforms all American open‑weight models, even models that are 3× larger,” and describes it as a “fully American coding model” because the supply chain for the base weights is U.S.‑based. The company is not releasing the weights publicly; instead, it plans a secure sharing program for trusted industry partners and research labs, according to unite.ai.

The launch follows Cognition’s earlier release of SWE-2 in September 2026, which was post‑trained on Moonshot’s Kimi K3 model. By contrast, Flash is the first Cognition coding model derived from a non‑Chinese open‑weight base, according to orcarouter.ai.

Sources

Fan coverage from Devin Central — not an official Cognition announcement. Devin is a trademark of Cognition.