transmission 05 · model wing, hall i
May 15, 2025 — weeks before the acquisition was signed — Windsurf shipped its own model family in the Wave 9 release: three tiers trained not to autocomplete code, but to survive the whole software-engineering process. It became the foundation of Cognition’s in-house model line.
The SWE-1 launch post argued that coding models were trained on the wrong task. Real engineering isn’t ‘produce the right snippet’ — it’s a shared timeline of half-finished states: the file you opened, the command that failed, the diff you’re about to revert. SWE-1 was trained with flow awareness — a model of the developer’s whole working session, not just the text cursor.
| tier | role | fine print |
|---|---|---|
| SWE-1 | the flagship tool-caller inside Cascade | ~Claude 3.5 Sonnet-level tool-call reasoning, cheaper to serve; free promo for paid users |
| SWE-1-lite | the unlimited tier | replaced Cascade Base at better quality; unlimited for every user, free or paid |
| SWE-1-mini | the reflexes | small enough to run inside Tab’s latency budget; further trained for predictive actions rather than tool calls |
We are launching our first family of models, dubbed SWE-1, optimized for the entire software engineering process, not just the task of coding. — Cognition, May 15, 2025
SWE-1 family (may 2025)
├── SWE-1 ....... flagship tool-call model, Cascade
├── SWE-1-lite .. unlimited tier, replaces Cascade Base
└── SWE-1-mini .. Tab predictive engine, latency-class
Next hall: the family gets a speed obsession — SWE-1.5.
— END OF TRANSMISSION · CTR-(ok)VERIFIED% —
+++ COGNITION ANNALS +++ DEVIN · THE AI SOFTWARE ENGINEER +++ SWE-1 → SWE-1.5 → SWE-1.6 → SWE-1.7 → SWE-2 +++ WINDSURF FOREVER +++ COMPILED BY DOGBED, NOT WRITTEN BY HAND +++ NO JAVASCRIPT WAS HARMED IN THE MAKING OF THIS ARCHIVE +++
rendered by dogbed — a document compiler, not a CMS. one binary, one template, thirteen documents.