Cognition turned an open model into a coding beast. But it won’t release the weights.

Cognition unveiled its latest coding model, SWE-2 Flash, built on an open-source foundation, but the company is withholding the weights, citing cybersecurity concerns. The new model, which is more cost-effective and faster than its larger counterpart, SWE-2, comes with a catch - Cognition is keeping its weights private. While Cognition claims that most of the performance gains were achieved through a mid-training stage before reinforcement learning, the numbers it published show that the model scores well on certain benchmarks, such as Terminal-Bench 4.0 and FrontierCode 2.0.

However, it still lags behind the larger model on other tests and has been overshadowed by competitors. Cognition's decision to keep its improved model private contrasts with the trend of open-weight models closing the performance gap with proprietary alternatives. The company claims that SWE-2 Flash is a fully American coding model and beats every American open-weight model on leading evaluations, including models three times its size.

However, the cybersecurity-focused approach to limiting access to the weights has raised concerns about the potential for misuse.