Cognition launches SWE-2 Flash, a coding model it says costs one-fifth as much as SWE-2 and runs 3x faster
Cognition has released SWE-2 Flash, a coding model available in Devin Desktop and CLI. The company says it costs one-fifth as much as SWE-2 and runs 3x as fast, and suits everyday work like code review. It is built on Thinking Machines' Inkling-Small, with 276 billion total and 12 billion active parameters. Cognition first further trained the base model on a few trillion tokens of scientific articles, blogs and code, then had it attempt many coding tasks and improve from the results. In Cognition's own tests, the base model scored 0.9% on Terminal-Bench 4.0 and SWE-2 Flash scored 20.0%; on FrontierCode 2.0 they scored 3.2% and 24.0%. Citing cyberattack risks, Cognition is withholding public weights and building a program to share them securely with trusted partners and labs.
Why it matters
Cheaper, faster coding models are taking over more routine engineering work, and Cognition is limiting weight access to trusted groups over cybersecurity risk.
Sources1
Written by dalao.ai: the title, summary and why it matters are our own words. The original reports are the record; their publishers keep authorship.