Cognition ships SWE-2 Flash in Devin, but the weights stay with partners

Cognition released SWE-2 Flash on November 4, a smaller and faster sibling of its SWE-2 coding model. It is available in Devin Desktop and the Devin CLI and nowhere else: there is no public API, and no SWE-2 Flash id in OpenRouter or models.dev as of 20:30 UTC.

WHAT COGNITION SAYS

SWE-2 Flash is post-trained from Inkling-Small, Thinking Machines' open-weight model with 276B total and 12B active parameters. For the first time, Cognition added a mid-training stage before reinforcement learning. It says SWE-2 Flash costs one-fifth of SWE-2 and runs about three times as fast, 200 tokens a second against 60.

On Cognition's own table, SWE-2 Flash scores 24.0 on FrontierCode 2.0 and 20.0 on Terminal-Bench 4.0. The base Inkling-Small scores 3.2 and 0.9 on the same rows, and the full SWE-2 scores 23.8 and 27.3. Cognition says it beats every American open-weight model on those evals, including models three times its size. These are Cognition's figures, not an independent run, and FrontierCode is Cognition's own benchmark.

THE WEIGHTS

The comparison is with open-weight models, but SWE-2 Flash's own weights are not public. Citing cyber risk, Cognition says it is building a program to share them securely with industry partners and researchers. There is no Hugging Face repository.

The same afternoon, Ricursive Intelligence released Ripple-65B Preview with downloadable weights under Apache 2.0, and it also claims the top US open-weight score on agentic coding. Cognition's table puts SWE-2 Flash at 89.2 on Terminal-Bench 2.1, and Ricursive reports 80.52 for Ripple. Each lab ran its own evaluation, so the two numbers are not directly comparable.

WHAT WE DO NOT KNOW

The per-token price, since Cognition gave only the ratio to SWE-2. Whether SWE-2 Flash will reach a public API. And who qualifies for the weights program.

Our SWE-2 Flash watch alerts the moment an id for it reaches a public API or a catalog we sweep.