SWE-2 Flash

SWE-2 Flash is a fast, lower-cost coding model designed to handle everyday software engineering tasks, including automation, code review, feature work, and documentation. Built on Inkling-Small, it adds a mid-training stage before reinforcement learning. Cognition reports that this raised performance on Terminal-Bench 4.0 from 0.9% to 20% and on FrontierCode 2.0 from 3.2% to 24%. The company says the model outperforms American open-weight competitors, including models up to three times larger. It is available in Devin Desktop and CLI at one-fifth the price and three times the speed of SWE-2, with generation speeds of about 200 tokens per second versus 60.

Training began with several trillion tokens of curated scientific writing, blogs, code, and other material. The team compared data mixtures using a held-out coding dataset, then trained in three stages, reserving the scarcest, highest-quality data for later. Reinforcement learning followed, with gains attributed mainly to more carefully checked tasks and a broader range of task types. The model’s ExploitBench score rose from 11.3% to 34.8%, reflecting added cybersecurity capability that could also enable misuse. Because of that risk, its weights are not publicly available. Cognition is developing a secure sharing program for trusted industry partners and research labs that need to inspect, fine-tune, or run the model in sensitive environments.