SWE-2 Flash: Cognition's new coding model for Devin runs 3x faster at one-fifth the price of SWE-2

Cognition introduced SWE-2 Flash on November 4, calling it a “fully American coding model” that sets a new bar for intelligence at low cost and high speed. It is available today in Devin Desktop and CLI, at one-fifth the price and three times the speed of SWE-2, Cognition’s larger coding model. Flash is free to use through November 30, 2026.

The pitch is deliberately narrow. Cognition says smaller models can now handle a significant share of the work software agents do — automations, code review, feature development, and updating documentation — and Flash is built for that everyday load rather than for pushing the capability frontier. The speed figure comes with a caveat worth noting: Cognition’s illustration replays each model’s own run on one Terminal-Bench task at simulated rates of 200 tokens per second for Flash and 60 for SWE-2, with transcripts shortened to equal length. It is a demonstration of the claimed gap, not a measured throughput benchmark.

Mid-training first, then reinforcement learning

Flash starts from Inkling-Small, Thinking Machines Lab’s mixture-of-experts model with 276 billion total parameters and 12 billion active parameters, whose full weights were released in July. Cognition added a large-scale mid-training stage before reinforcement learning — the first time it has used that order in its recipe. To pick the data mix, the team built a held-out coding dataset meant to reflect real software-engineering work, then compared candidate mixtures by how many tokens each needed to reach the same validation loss as a baseline run.

The final mix ran to a few trillion tokens drawn from scientific articles, blogs, code, and other sources, taught in three stages so the scarcest, highest-quality data came last. Cognition reports held-out validation loss improving at a stable log-linear rate through those stages. Large-scale reinforcement learning followed, with most of the gains over the SWE-2 recipe attributed to higher-quality, more diverse tasks — instructions, environment setups, and verifiers audited by the team, plus new task types covering a wider range of the reasoning Devin work demands.

Faster and cheaper — but not better at everything

On Cognition’s reported figures, the training paid off against the base model and mostly held its own against SWE-2. Terminal-Bench 4.0 rose from 0.9% for Inkling-Small to 20.0% for Flash, and FrontierCode 2.0 from 3.2% to 24.0%. Against SWE-2, Flash edges ahead on FrontierCode 2.0 (24.0% versus 23.8%) but trails on Terminal-Bench 4.0 (20.0% versus 27.3%) and on DeepSWE 1.1 (60.1% versus 73.0%). Cognition also claims Flash outperforms all American open-weight models on the evaluations it reports, including models three times larger — its claim, on its chosen evaluations, and the figures are vendor-reported.

That framing matters. On Cognition’s own numbers, the larger model still leads on long-horizon terminal tasks, while Flash’s case rests on volume: routine jobs at a fifth of the price. It is a cost-and-speed option for everyday agent work, not an across-the-board replacement for SWE-2.

The weights stay locked up — for now

The most consequential decision in the launch is what Cognition did not release. Targeted cybersecurity training lifted Flash’s ExploitBench score from 11.3% to 34.8%, and Cognition says those same capabilities could assist malicious attackers, pointing to a recent surge in cyberattacks discussed in its October security post. The weights are therefore not being shared publicly yet. Instead, Cognition is building a program to share them securely with trusted industry partners and research labs, with more detail promised in a few weeks.

The program is aimed at the customers who need more than an API: enterprises and governments that must host and control deployments on their own infrastructure, including on-premises, and researchers who need to study how the model behaves and how it can be misused. Approved partners would be able to fine-tune the model and run it in sensitive, customer-controlled environments. Until that program exists, Flash is a hosted model inside Devin — useful, but not inspectable in the way its open-weight base model is.

The open question

Flash joins a crowded field of coding models built for long engineering work, including Meituan’s LongCat-2.5. The contrast with open weights is sharpest elsewhere in today’s news: Kimi K3.1 Pro’s weights landed on Hugging Face the same day, while Flash’s stay private.

The checkable facts are the price, the speed claim, and the benchmark table — all Cognition’s. The free period through November 30 in Devin Desktop and CLI is the low-risk way to test whether the smaller model handles a team’s routine work. Whether the secure-sharing program delivers real weight access to partners and researchers is the question the next few weeks are meant to answer.