A smaller coding model earns its place when the patch survives review, not when its tokens arrive first.
Cognition announced SWE-2 Flash on November 4 for Devin Desktop and Devin CLI. Its launch page advertises free use through November 30. The company also says it is preparing controlled weight access, while keeping the weights out of public distribution for now.
The hosted model can be tried before weights are shared
Flash builds on Thinking Machines’ Inkling-Small, according to Cognition. The base has 276 billion total parameters and 12 billion active parameters. Cognition added mid-training and reinforcement learning for software-agent work.
Cognition claims one-fifth the cost and three times the speed of the larger SWE-2. Its animated speed comparison replays shortened transcripts at simulated rates. That illustration is not an independent latency measurement or a published token-price schedule.
The benchmark table does not show a uniform win
The provider’s table gives Flash a slight lead over SWE-2 on FrontierCode 2.0: 24.0 against 23.8. On DeepSWE 1.1, Flash scores 60.1 against 73.0. Terminal-Bench 4.0 also favors SWE-2, at 27.3 against Flash’s 20.0. All are Cognition-reported results.
Those mixed results make task selection useful. A team could evaluate Flash on small documentation edits, code review, or a bounded fix before assigning a larger change. Our SWE-2 coverage explains the earlier model and its evaluation context.
Judge the cost of an accepted change
For a comparison, give both models the same repository state and issue. Keep tests that expose the original defect, then record whether the final patch passes them. Include human repair time and rejected attempts in the result. This is a proposed evaluation; we have not run it.
A fivefold reduction in inference cost only carries through to total spend if usage and accepted-output rates remain comparable. A model that needs substantially more retries can reduce the savings. The temporary offer provides a trial window, but you still need to check account entitlements and usage limits inside Devin.
Self-hosting is a later decision. Cognition says a secure sharing program for partners and research labs is being built; it has not announced public weights. Which recurring repository task would you use to compare Flash with your current coding model?
Abdessalam Alaoui
I'm Abdessalam Alaoui, founder of Musthave.AI. I write practical guides to AI models, tools, generative media.
