Cognition built SWE-2 Flash on an open base, kept weights private

Cognition built its new coding model on a model anyone can download. It then kept the finished weights for itself. The company published the decision on Wednesday with the launch of SWE-2 Flash.

Cognition says SWE-2 Flash costs one-fifth as much as its larger SWE-2 model. The smaller model generates 200 tokens per second against SWE-2's 60, according to Cognition. SWE-2 Flash sits inside Devin Desktop and the Devin CLI. It is free to try until 30 November.

What the model is built on

The base is Inkling-Small from Thinking Machines Lab. Inkling-Small is a 276-billion-parameter mixture-of-experts model with 12 billion active parameters. Thinking Machines released it on 30 July under the Apache 2.0 licence. That licence lets anyone pull the weights from Hugging Face and build on them. It does not force a company to share its own work afterwards.

Cognition says the biggest gain came from adding a mid-training stage before reinforcement learning. The step used "a few trillion tokens of high-quality scientific articles, blogs, code, and other sources", the company says. Cognition's own numbers show the step working. On Terminal-Bench 4.0, the finished model scored 20.0 percent. Inkling-Small scored 0.9 percent. On FrontierCode 2.0, the pair scored 24.0 percent and 3.2 percent. On DeepSWE 1.1, 60.1 percent and 17.4 percent.

Why the weights stay closed

Cognition says cybersecurity drove the decision. The company put cybersecurity work into Flash's training. ExploitBench moved from Inkling-Small's 11.3 percent to Flash's 34.8 percent, Cognition says. These capabilities help find and investigate flaws, the company adds. They can also help attackers.

The benchmark does not say whether a public release would raise real-world risk. Cognition is building a programme to share the weights with picked industry partners and research labs. Those partners could inspect, fine-tune and run the model on their own hardware. The company says more details are coming in the next few weeks. For now, it is not public who will qualify or what the licence will say.

What the vendor claims and what to check

Cognition calls SWE-2 Flash "a fully American coding model" and says it beats every American open-weight model on leading evaluations. Cognition's own benchmark table shows models built outside the United States scoring higher than Flash on several tests. DeepSeek V4.1 Flash is one of them.

Cognition writes that Flash nearly matches the larger SWE-2 on SWE-Bench Pro v2 Hard. The scores are 89.9 percent and 90.6 percent. Flash trails SWE-2 on DeepSWE 1.1 and Terminal-Bench 4.0. Claude Haiku 5.5 and DeepSeek V4.1 Flash outscore Flash on several benchmarks in the same table.

Every one of these numbers is Cognition's own. The model has not been through an independent evaluator since Wednesday. A buyer testing Flash in Devin is seeing a run against the vendor's chosen cases.

A team weighing a coding model this month now has two shapes of answer. One is an open-weights release the team runs inside its own security boundary. The other is a hosted model with the base open and the final layer sealed. Cognition is offering the second. The market will decide if the cybersecurity reason is enough to keep buying it.

Source

Cognition's SWE-2 Flash post. Reported in The New Stack by Amanda Caswell.