AI modelsNovember 04, 2026
-31
🧊
Cognition's SWE-2 Flash Shocks Coding World
Cognition ships SWE-2 Flash, a coding model refined through mid-training that is three times faster and available at one-fifth the price.
#coding model#Cognition#SWE-2#reinforcement learning#Devin
🔥 What happened
Cognition dropped SWE-2 Flash, a coding model that jumps Terminal-Bench 4.0 from 0.9% to 20.0% — a 22x leap. It's 5x cheaper and 3x faster than its bigger sibling SWE-2. Weights stay closed for now, with a gated program for select partners.
💡 Why it matters
A 276B-parameter model with just 12B active parameters beats American open-weight models 3x its size. The secret sauce: a new mid-training stage before RL that massively boosts token efficiency. For agent pipelines, that means 200 tok/s instead of 60 — tasks that took minutes now finish in seconds.
⚡ Our take
Cognition proves the future isn't bigger models, it's smarter training recipes. Keeping the weights closed is a tactical mistake — the community will reverse-engineer it anyway.
The title, summary and analysis of this item were produced automatically by an AI system and have not been editorially reviewed. They may contain errors, bias or omissions — when in doubt, read the linked original source.