By No Code MBA · Facts checked
- Provider
- Cognition
- Announcement date
- Status at last check
- Generally available
What changed
Cognition describes SWE-2 Flash as a coding model built by adding a mid-training stage before reinforcement learning on top of Inkling-Small. Cognition reports Terminal-Bench 4.0 results of 20.0% versus 0.9% for the base model and FrontierCode 2.0 results of 24.0% versus 3.2%; these are the company's evaluations.
Access and availability
Available in Devin Desktop and CLI. Cognition says the model weights are not publicly shared and that it is building a program to provide them securely to trusted industry partners and researchers.
Official access documentationWhat you can build
Try it on a bounded coding task in Devin and compare the output, speed, and review effort with your current model before expanding its role.
Limitations
Cognition says it is not sharing weights publicly because of cybersecurity risks; the secure weight-sharing program is still being built. Benchmark and performance claims on the announcement are Cognition's own.