Cognition launched SWE-2 Flash on Tuesday, November 4, 2026, a smaller and cheaper version of its SWE-2 coding model, aimed at everyday work inside the Devin agent. According to the company, the new model runs 3 times faster and costs less than one-fifth of SWE-2 — a combination designed for routine engineering tasks such as automations, code review, feature development and documentation updates.
This is an update to Geekiki's coverage of the SWE-2 line: in September, Cognition introduced the original SWE-2, promising near-frontier performance at up to 70% lower cost than rivals. Read the earlier article: Cognition launches SWE-2: a coding model near the frontier and up to 70% cheaper.
A smaller model, built on Inkling-Small
SWE-2 Flash has 276 billion total parameters and 12 billion active per token (276B-A12B). It starts from Inkling-Small, Thinking Machines' mixture-of-experts base model, and received an extra mid-training stage before reinforcement learning — a step Cognition says was decisive in lifting a small checkpoint to reliable coding-agent level. For comparison, the original SWE-2 has 2.8 trillion parameters, with 104 billion active.
On inference speed, Cognition cites around 200 tokens per second for Flash, versus around 60 for SWE-2, in the company's own benchmark replay — a lab figure, not a live race between the two models.
Benchmarks: a leap over the base, a near tie with its bigger sibling
All figures below were published by Cognition itself and have no independent evaluation yet. On Terminal-Bench 4.0, Flash scores 20.0%, a huge jump from the base model's 0.9%, but below SWE-2's 27.3%. On FrontierCode 2.0, it reaches 24.0% — slightly above SWE-2's 23.8%. On SWE-Bench Pro v2 Hard, it records 89.9%, nearly tied with the larger model's 90.6%.
The most sensitive point is ExploitBench, an offensive-security benchmark: Flash jumps from 11.3% (base) to 34.8%, close to SWE-2's 36.1%. That gain is behind the launch's most controversial decision.
Weights withheld: offensive capability comes at a price
Unlike the base model — Inkling-Small is open, under the Apache 2.0 licence — Cognition will not publicly share SWE-2 Flash's weights. The company argues that the same capabilities that help cybersecurity defenders can also be used by attackers to create new cyberattacks and scale existing ones, at a time of rising cyberattacks. Instead, it promises a secure weight-sharing programme for trusted industry partners and research labs, with details in the coming weeks.
In practice, Flash cannot currently be downloaded or run outside Devin — a notable difference from the same week's wave of open models, such as Baidu's ERNIE 5.5 Flash Base and Ricursive's Ripple-65B.
Availability
SWE-2 Flash is already available in Devin Desktop and the Devin CLI, with access also via the web and mobile app, and is free through November 30, 2026. Cognition has not disclosed a per-token price, context window or a full model card.
Image credit: Devin logo, published in Cognition's official brand kit.
Sources
Cognition — Introducing SWE-2 Flash
Unite.AI — Cognition Launches SWE-2 Flash for Faster, Lower-Cost Coding Agents
Superpower Daily — Cognition adds a lower-priced coding model to Devin, keeping its weights private
Geekiki — Cognition lança SWE-2: modelo de código perto da fronteira e até 70% mais barato
Transparency: this article was produced with the assistance of artificial intelligence, cross-checking public posts on X and sources on the internet. Consult the original sources to verify the details before making any decision.
Muse
