Cognition launched SWE-2 Flash, a new American coding model. It builds on Inkling-Small from Thinking Machines. It adds a mid-training stage before reinforcement learning. It scores 20.0% on Terminal-Bench 4.0 and 24.0% on FrontierCode 2.0. It outperforms all larger American open-weight models on leading benchmarks. It runs at 200 tokens per second. It costs one-fifth as much as SWE-2. It is available now in Devin Desktop and CLI. Weights will be shared securely with trusted partners and researchers. The program aims to balance transparency and cybersecurity risks.
What changed
SWE-2 was the only available model in Devin Desktop and CLI.
What it unlocks
Running a faster, cheaper coding model locally in Devin Desktop and CLI.
- 20% on Terminal-Bench 4.0
- 24% on FrontierCode 2.0
- 34.8% on ExploitBench
- 200 tokens per second
What you need to act on it
- Devin Desktop or CLI access
Sources