Cognition's SWE-2 Flash Undercuts Its Own Big Model

Cognition, the company behind the Devin coding agent, released a lighter model on Tuesday called SWE-2 Flash. The pitch is speed and price. Cognition says it runs about three times faster than its larger SWE-2 model and costs roughly a fifth as much, and it is free for paying Devin subscribers through November 30.

Under the hood, Flash is a mixture-of-experts model with 276 billion total parameters but only 12 billion active at any time, which is what keeps it quick and cheap to run. It is built on Inkling-Small, a base model from Thinking Machines Lab whose weights were published in July. For now Flash works inside Devin Desktop and the Devin command-line tool, with no public API, model identifier or per-token price announced yet.

The benchmark picture is mixed, to Cognition's credit for saying so. Flash edges past the full SWE-2 on one test, FrontierCode 2.0, while SWE-2 still scores higher on Terminal-Bench 4.0 and DeepSWE 1.1. Cognition positions the small model for the bread-and-butter work of a coding agent, things like automations, code review, routine features and documentation, and keeps SWE-2 for the harder problems. The company also claims Flash beats every American open-weight model on the evaluations it reports.

That last boast carries an edge, because Cognition is not releasing Flash's weights. It cited cybersecurity risk and said it would share the checkpoint only with selected partners and researchers through a vetted program. A model trained on open foundations, then closed back up on the way out the door, is a neat illustration of where the industry's instincts are pulling.

The release fits a clear direction of travel: cheaper, faster tiers aimed at the everyday tasks that make up most of an agent's day. OpenAI did much the same when it made computer use cheaper with GPT-6.1 Terra, and a wave of smaller open-weight coding models has been chasing the same ground. The twist here is the lock on the door, at a time when some open-source projects have started turning AI-written code away altogether.