Cognition, the company behind the Devin coding agent, released SWE-2 Flash on November 4: a smaller version of its SWE-2 coding model that it says runs about three times faster at one-fifth the price. It is available now in Devin Desktop and the Devin CLI, and Cognition says paid Devin users can use it free through November 30.
Why it matters: most of what coding agents do all day is not hard. Code review, documentation updates and routine automations burn tokens on long loops where speed and cost matter more than peak reasoning. A cheaper tier inside the same agent lets teams route that work away from the expensive model. It is also another sign that "Flash"-style mid-size models built on open bases are becoming the default workhorse layer for agents.
What's under the hood
According to Unite.AI's summary of Cognition's announcement, SWE-2 Flash is a mixture-of-experts model with 276 billion total parameters and 12 billion active, adapted from Inkling-Small, an open-weight base model from Thinking Machines Lab. RuntimeWire describes the base as a small American open-weight model that Cognition post-trained.
The speed claim comes from a simulation: Cognition replayed Terminal-Bench 4.0 runs, with transcripts trimmed to equal length, showing Flash at about 200 tokens per second versus about 60 for SWE-2, per Unite.AI.
The benchmarks Cognition reported
These are Cognition's own numbers as reported by Unite.AI; we haven't seen independent runs.
| Evaluation | Inkling-Small (base) | SWE-2 Flash | SWE-2 |
|---|---|---|---|
| Terminal-Bench 4.0 | 0.9 | 20.0 | 27.3 |
| FrontierCode 2.0 | 3.2 | 24.0 | 23.8 |
| DeepSWE 1.1 | 17.4 | 60.1 | 73.0 |
Two things stand out. First, Cognition's post-training does most of the work: the base model barely registers on Terminal-Bench, while Flash scores 20. Second, Flash only edges SWE-2 on FrontierCode; on DeepSWE it trails by about 13 points and on Terminal-Bench by about 7. That is the trade you make for the speed and price.
Cognition also claims Flash beats every American open-weight model on the evaluations it reports, according to both outlets. Note that this compares a model whose weights are not public against models whose weights are.
Weights stay gated
SWE-2 Flash is not an open release. Both outlets report that public weights are not available; Cognition plans a restricted sharing program for trusted partners and research labs, with details to come. Unite.AI ties that caution to cybersecurity training that Cognition says lifted the model's ExploitBench score from 11.3% to 34.8%.
What's still missing
Neither report gives absolute per-token prices, only the one-fifth ratio, and neither says what Flash will cost in Devin after November 30. There is also no word on API access outside Devin.
Who should care
- Devin teams: the free period through November 30 is a cheap way to test whether Flash is good enough for reviews, docs and routine automations before you pay for it.
- Anyone picking a coding model: treat the comparison table as a vendor claim. The gap on DeepSWE suggests keeping the larger model for harder, multi-step fixes.
- Open-model watchers: a strong result built on Inkling-Small is a data point for Thinking Machines' base, even if Cognition's fine-tuned weights stay closed.
