Cognition says SWE-2 Flash costs one-fifth as much as SWE-2 and runs about 3x faster (their demo replays 200 vs 60 tokens per second). It starts from Inkling-Small, a 276B mixture-of-experts model with 12B active parameters, and adds a mid-training stage of a few trillion tokens before reinforcement learning.
The headline gains are over the base model: Terminal-Bench 4.0 went from 0.9% to 20.0% and FrontierCode 2.0 from 3.2% to 24.0%, per Cognition's table. Against its big sibling it's close on some evals (Terminal-Bench 2.1 is 89.2 vs 92.8) and further back on others (DeepSWE 1.1 is 60.1 vs 73.0). All of these are Cognition's own numbers.
Cognition also trained it on security work, lifting ExploitBench from 11.3% to 34.8%. Because of that, it is holding the weights back and says it is building a program to share them with vetted partners and labs "in a few weeks".