Cognition built a better coder on open weights, then kept its own gated

On November 4, Cognition put SWE-2 Flash into Devin Desktop and CLI. The pitch is simple: three times the speed and a fifth of the price of SWE-2, free through November 30, aimed at the routine work agents do all day, automations, code review, feature work, documentation. A solid pitch, and the reported benchmarks mostly back it. The decision that will outlast the pricing page sits near the bottom of the announcement: the weights are not coming out.

Flash carries 276 billion total parameters with 12 billion active per token, adapted from Inkling-Small, the mixture-of-experts base Thinking Machines released with open weights on July 30. Cognition added a large-scale mid-training stage before reinforcement learning, a first for its recipe, and picked the data mix with token-efficiency plots: whichever candidate mixture reached the baseline validation loss on fewer tokens won. A few trillion tokens went in, staged so the scarcest, highest-quality data arrived last, and held-out validation loss improved at a stable log-linear rate through all three stages.

The gains over the base are large. Terminal-Bench 4.0 went from 0.9 to 20.0 percent. FrontierCode 2.0 went from 3.2 to 24.0, which is also the one eval where Flash edges SWE-2, at 24.0 versus 23.8. DeepSWE 1.1 went from 17.4 to 60.1, still well under SWE-2's 73.0. Cognition also states that Flash outperforms every American open-weight model on the evals it reports, including models three times larger, which reads a little odd in a release about a model whose own weights are gated. The speed comparison is a replay, each model's own Terminal-Bench run simulated at 200 tokens per second against 60, with transcripts shortened to equal length. Directionally informative, not a stopwatch.

The number that gated the release

The security paragraph is doing the gating. Targeted cybersecurity training raised ExploitBench from 11.3 to 34.8 percent, and Cognition's own reasoning is that the same capability serves attackers, a point its October security post ties to a surge in real attacks. So the weights stay inside the company. The planned outlet is a secure sharing program for vetted partners and research labs, details promised in a few weeks, with fine-tuning and on-premise deployment as the incentive to join. What exists today is a model you can use and cannot hold.

Same fork, opposite exits

Two releases in two weeks hit the same fork and took opposite sides. Moonshot measured a climbing cyber capability in Kimi K3.1 Pro's training runs, stripped the cyber data out of the pipeline, and shipped full open weights anyway (covered here on November 2). Cognition built the cyber capability in, kept it, and withheld the weights pending a vetted-access program. Same measurement tooling, opposite exits, and both companies published their numbers, which makes the comparison unusually clean. What used to be a default, weights downloadable within the week, has quietly turned into a per-release judgment call.

What you can run today

The practical path is Devin. Flash is live in Desktop and CLI, free until November 30, positioned for exactly the high-volume, low-glamour work that dominates agent bills. If your workflow needs local weights, the base is still downloadable: Inkling-Small went out open and sits on Hugging Face. What you will not get is the reinforcement learning that turned it into a coding model, none of the task mix that moved Terminal-Bench nineteen points. Fine-tuning the base yourself buys a starting position, not the product.

Model releases used to end with a license file. They now end with an access decision, argued in the announcement itself, and Flash at least shows its work: the same training that made the model more useful to defenders made it more useful to attackers, and the company picked a side. The base is on Hugging Face, the build on top is not. Watch whether the partner program ships in a few weeks as promised, because whichever exit lands first becomes the one every later release gets compared against.