From the source
Cognition introduced SWE-2 Flash, a coding model it says outperforms all American open-weight models, including ones three times larger, and made it available in Devin Desktop and CLI.
- The model is offered at one-fifth the price and three times the speed of its larger sibling, SWE-2, aimed at everyday engineering work such as automations, code review, feature development and documentation updates.
- Cognition says it added a mid-training stage before reinforcement learning for the first time, which lifted Terminal-Bench 4.0 from 0.9% to 20.0% and FrontierCode 2.0 from 3.2% to 24.0% over the base model, Inkling-Small from Thinking Machines.
- Mid-training ran on a few trillion tokens of scientific articles, blogs and code, split into three stages so the scarcest, highest-quality data came later, followed by large-scale reinforcement learning with audited tasks and new task types.
- Cognition says it targeted cybersecurity work in training, raising ExploitBench from 11.3% to 34.8%, and is not sharing the weights publicly yet because the same capabilities could aid attackers.
- Instead it is building a program to securely share the weights with trusted industry partners and research labs, with more details promised in a few weeks.
