Introducing SWE-2 Flash: Eight Benchmark Rows and Not One Endpoint

Cognition announced SWE-2 Flash on 4 November 2026, and the post is unusually generous with numbers: eight benchmark rows, two before-and-after training deltas, a speed figure and a price figure. What it does not contain is a way to call the model. SWE-2 Flash runs inside Devin Desktop and the Devin CLI, free through 30 November 2026. There is no API endpoint, no model ID, no per-token rate and no public weights. For a launch whose headline is "fast and efficient", the most useful thing a developer can know is that the efficiency is currently measured in dollars you do not pay rather than tokens you can buy.

That is not a knock on the model. It is the shape of the offer, and it decides who can act on this announcement today and who has to wait.

What actually shipped on 4 November

The spec sheet is short enough to hold in one screen, which is itself informative — Cognition published less about this model than it published about SWE-2 in September.

• Base model — post-trained from Inkling-Small, Thinking Machines Lab's open-weight mixture-of-experts at 276B total parameters and 12B active, Apache 2.0 upstream.

• Serving — Devin Desktop and the Devin CLI only. No hosted endpoint, no vendor API, no weights download.

• Price — no per-token rate is published. Access is free through 30 November 2026.

• Speed — roughly 200 tokens per second against SWE-2's 60, a figure Cognition produced by replaying shortened session transcripts at simulated rates rather than timing a live endpoint.

• Context window, maximum output, and effort settings — not stated anywhere in the launch post.

The base-model swap is the genuinely new engineering. SWE-2 was built on Kimi K3, a 2.8T-total / 104B-active base. SWE-2 Flash sits on a base roughly a tenth the size by total parameters and an eighth by active parameters, and Cognition says most of the quality recovery came from better reinforcement-learning tasks and audited verifiers rather than from scale. The post also describes the first mid-training stage the company has used before RL: a few trillion tokens of scientific articles, blogs and code, run in three passes with the rarest data last.

The claim that will be repeated, and the row it rests on

The sentence everyone will quote is that SWE-2 Flash "outperforms all American open-weight models, even ones 3x larger". It is worth knowing exactly how much of that survives contact with the table underneath it.

The only direct, same-benchmark comparison in the launch material against a smaller American open-weight coder is Terminal-Bench 2.1, where SWE-2 Flash reports 89.2 against Flint v1.0's 54.9. That is a large gap on a shared benchmark version, and it is the strongest single piece of evidence for the claim. But Flint v1.0 is 39.8B total, and Cohere's North Mini Code is 30B — neither is 3x larger than a 276B model. They are smaller, which makes beating them unremarkable rather than impressive.

The 3x-larger comparators in Cognition's own table are Muse Spark Open at 616B total and DeepSeek V4.1 Flash 1017 at 763B total. On DeepSWE 1.1, SWE-2 Flash reports 60.1 against Muse Spark Open's 60.0 — a tenth of a point, inside any plausible noise band, and the model it is being compared to is not described as open-weight in the same breath. DeepSeek V4.1 Flash 1017, meanwhile, is not American, so it cannot be part of an "American open-weight" claim at all. It is also the model that beats SWE-2 Flash on Terminal-Bench 4.0 by 22.3 points — 42.3 to 20.0 — which is the largest single loss anywhere in the comparison.

None of this makes the claim false. It makes the claim narrower than it sounds: one shared benchmark against one American open-weight model, and a rounding-error win against a model whose licence is a separate question. All of these figures are Cognition's own, run on a harness the post does not name.

Two Terminal-Bench numbers, seventy points apart

If you read only one thing off Cognition's table, read this pair. SWE-2 Flash scores 20.0 on Terminal-Bench 4.0 and 89.2 on Terminal-Bench 2.1. Those are not competing estimates of the same thing; they are different benchmark generations with different task sets and different difficulty, and the version number is the only thing separating them at a glance. A launch table that lists both invites exactly the error of comparing 20.0 against a rival's 2.1 score, or 89.2 against a rival's 4.0 score.

The safe rule for this model, and for any model whose vendor reports more than one Terminal-Bench version: compare 2.1 to 2.1 and 4.0 to 4.0, and say which you are quoting. We keep a running note on this because the launch table's flattened text rendering has already misassigned one row in secondary coverage.

Why the weights are behind a door — and what the door is

Cognition frames the withheld weights as a transparency-versus-risk trade, and the post is unusually direct about the risk side. The company says it deliberately trained for security work and lifted its ExploitBench score from 11.3 during base training to 34.8 after post-training. Publishing those weights would put a competent exploit-writing model in general circulation.

So instead there is a programme. Weights will be shared "securely" with partners and researchers, and the post promises details "in a few weeks". As of this writing there is no application form, no eligibility criteria, no licence, no date and no statement about whether shared weights may be redistributed or used commercially. A partner programme is not open weights, and the distinction matters for anyone planning to self-host: until the terms exist, the correct planning assumption is that SWE-2 Flash will never be downloadable.

Free until 30 November, then unpriced

The commercial offer is a month of free access inside Devin. After that, nobody outside Cognition knows what it costs, because no per-token rate was published. The "one fifth the price of SWE-2" line is a ratio against a product with its own bundled pricing, not a rate card.

That is a real planning risk if you are considering building a workflow around it. An unpriced dependency is worse than an expensive one, because you cannot model it. The sensible move between now and 30 November is to measure the thing you can measure — whether the model's output quality in your own repository justifies switching your Devin configuration — and to treat the cost question as open until Cognition publishes a number.

What you can actually call today

SWE-2 Flash is not on OrcaRouter and will not be until Cognition ships an endpoint or a licensed weight release; we are not going to pretend otherwise. What is callable right now is the rest of the field this launch is implicitly competing with — Meta's Muse Spark 1.2 and 1.1 among them, both of which are live in the catalogue and both of which are the models that set SWE-2 Flash's DeepSWE 1.1 score to a tenth of a point.

If the reason you are reading this is that you want a cheap, fast coding model in an agent loop this week rather than in December, that is the practical fork: wait for Cognition's pricing and partner programme, or point your existing client at one API for more than 200 models and A/B the alternatives against your own test suite. OrcaRouter passes provider list price through at 0% markup, so if Cognition eventually publishes a rate, that rate is what you would see here on day one rather than a marked-up reseller number.

What the launch post left out

Four omissions are worth flagging because they are the ones that decide whether the numbers mean anything:

• No harness. The September SWE-2 post disclosed its scaffold and effort setting. This one does not, which makes every agentic score a black box.

• No best-of-N disclosure. We do not know whether these are pass@1 or a best-of-five.

• No context window. For a model positioned as the cheap option inside an agent, the context budget is not a footnote.

• No independent run. Every figure here is vendor-reported and unreproduced, including the FrontierCode 2.0 score of 24.0 — and FrontierCode is Cognition's own benchmark, which is worth remembering when it is the row that shows a win.

The thing to watch is not another benchmark row. It is the partner programme page: the date, the licence and the price. When those three exist, SWE-2 Flash becomes a model you can evaluate. Until then it is a well-documented demo with a free month attached.