Cognition releases SWE-2 Flash, free in Devin Desktop and CLI through Nov 30

On November 4, 2026, Cognition released SWE-2 Flash, a 276B-parameter mixture-of-experts coding model with 12B active parameters, free in Devin Desktop and Devin CLI through November 30.

Archive entry: this covers an event from November 4, 2026, and we published it on November 8, 2026. The details come from sources dated at the time and may have changed since.

On November 4, 2026, Cognition released SWE-2 Flash, a mixture-of-experts coding model, as described in Cognition's announcement post. The model was available in Devin Desktop and Devin CLI and was free through November 30.

What Cognition announced

SWE-2 Flash was a 276B-A12B model, with 276B total and 12B active parameters, built on Inkling-Small from Thinking Machines.
Cognition called it a fully American coding model and claimed it beat every US-made open-weight model on major evaluations.
Training added a mid-training stage on a few trillion tokens before reinforcement learning that reused the SWE-2 recipe.
Reported scores were 24.0 on FrontierCode 2.0 (SWE-2: 23.8), 45.5 on FrontierCode 1.1 Main (SWE-2: 50.0), 60.1 on DeepSWE 1.1 (SWE-2: 73.0) and 89.2 on Terminal-Bench 2.1 (SWE-2: 92.8).
Speed was 200 tokens per second versus 60 for SWE-2, measured in a simulated replay of one Terminal-Bench 4.0 task.
Cognition said the model cost one-fifth the price of SWE-2 and gave no per-token rate.

What it meant for people building with AI

Devin Desktop and Devin CLI users could try a fast, cheaper in-house coding model at no cost through the end of November. Against SWE-2, the reported scores were higher on FrontierCode 2.0 and lower on the other three benchmarks, and the speed gain was measured in a single simulated replay.

Weights and next steps

Weights were not public at the time of the announcement. Cognition said it planned a program to share them securely with vetted partners and researchers, with details promised in a few weeks.

Sources

Introducing SWE-2 Flash (Cognition, November 4, 2026, official)

Written with AI assistance, strictly from the sources listed above. Before it went live, automated checks verified its dates, sources and every language edition.