Agent Upstart

Cool things people are building with agents — and practical help getting the most out of them.

Playbooks · Coding tasks

SWE-2 Flash: Devin's cheaper model, closed weights

Written November 5, 2026

Cognition's SWE-2 Flash runs at a fifth of the SWE-2 price and three times the speed, free through November 30. The weights stay private.

What it is

Cognition introduced SWE-2 Flash on November 4, a smaller coding model for Devin Desktop and Devin CLI. It is built from Thinking Machines' Inkling-Small, a sparse mixture-of-experts base model with 276 billion total parameters and 12 billion active per token. Cognition added a mid-training stage before reinforcement learning, then post-trained it for everyday engineering work. The announcement puts it plainly: one-fifth the price and three times the speed of its much larger sibling, SWE-2, free through November 30 in Devin Desktop and CLI.

The price and speed figures are relative, not absolute. Cognition has not published standalone token rates for Flash, so "one-fifth the price" is a comparison against SWE-2, not a rate to put in a spreadsheet. The free period is a promotion with an end date: after November 30, Flash moves to paid use, and the rate then is not yet published. Check the in-app model picker before budgeting past the free window.

The trade, in Cognition's own numbers

All scores below are Cognition-reported, and the comparison figures come from Cognition's announcement and its coverage. Flash is not a straight upgrade over SWE-2; it wins some tests and loses others. On FrontierCode 2.0, Flash scores 24.0% against 23.8% for SWE-2 — a narrow win. On Terminal-Bench 4.0, Flash scores 20.0% against 27.3% for SWE-2 — it trails by a clear margin on the test closest to agentic terminal work. On SWE-Bench Pro v2 Hard, Flash scores 89.9% against 90.6% for SWE-2 — near enough to a tie. Against its own base model the improvement is large: mid-training and reinforcement learning took Terminal-Bench 4.0 from 0.9% on Inkling-Small to 20.0% on Flash, and FrontierCode 2.0 from 3.2% to 24.0%.

The speed claim needs the same care. Cognition's generation-speed illustration replays each model's own Terminal-Bench 4.0 run at simulated rates of 200 tokens per second for Flash and 60 for SWE-2, with transcripts shortened to equal length. That is an illustration of the claimed speed ratio, not a measured head-to-head run.

The weights stay private

Flash's weights are not being shared publicly. Cognition cites cybersecurity risk: it targeted security training in the model, raising its ExploitBench score from 11.3% to 34.8%, and says those same capabilities could assist malicious attackers. It is building a program to share the weights securely with trusted industry partners and research labs, with more detail promised in a few weeks. The public does not get access.

Do not confuse the base model with Flash. Inkling-Small, the base model Cognition started from, is downloadable under the Apache 2.0 licence. That availability is not a Flash release: the post-training that makes Flash what it is stays inside Cognition, and downloading Inkling-Small does not give the Flash model.

What it fits, and what stays on SWE-2

Cognition positions Flash for routine work: automations, code review, feature development, and documentation updates. That is where a cheaper, faster model earns its keep — high-volume, lower-stakes tasks where the per-task cost matters more than the last few points of capability. Harder work stays on SWE-2, where the Terminal-Bench gap is the warning: a model that trails by seven points on terminal-based agent work will cost more in retries and failed runs on tasks that lean on that skill. The sensible split is to route routine work to Flash while it is free, measure cost per finished task rather than cost per token, and keep SWE-2 for work where a failed first attempt is expensive.

Limits

Related playbooks

Sources

Back to: Coding playbooks hub · Also: GPT-6.1 Sol, Ultrafast, and the new Pro tiers · Factory Droid 1.0 and the harness-efficiency claim