ModelsHigh importance

Cognition launches SWE-2 Flash (276B-A12B): 1/5 the price and 3× the speed of SWE-2, weights still not public

Cognition unveiled (11/04) SWE-2 Flash, a US coding model at 276B-A12B. For the first time the recipe includes mid-training before RL, starting from Inkling-Small (Thinking Machines): Terminal-Bench 4.0 rises from 0.9% to 20.0% and FrontierCode 2.0 from 3.2% to 24.0%. On leading evals, it beats US open-weight models even ~3× larger. On Devin Desktop and CLI, it costs 1/5 and runs at 3× the speed of SWE-2 (about 200 vs 60 tok/s). ExploitBench rises from 11.3% to 34.8%. Weights are still not public; Cognition is building a secure sharing program with partners and researchers.