Cognition Launches SWE-2 Flash: A Fast Coding Model Built on Inkling-Small, 200 tok/s at One-Fifth the Price of SWE-2

Background and the Industry Pain Point

Much of what coding agents do every day, such as automations, code review, small features and documentation updates, does not need the largest frontier model. Cognition says that alongside pushing the frontier it wants to cut the cost of everyday engineering work, which is why it built the smaller and faster SWE-2 Flash.

Architecture Highlights and Mechanism

SWE-2 Flash starts from Thinking Machines' Inkling-Small (276B total, 12B active). For the first time Cognition added a large mid-training stage before RL: it chose the data mix using token-efficiency curves, trained in three stages on a few trillion tokens of scientific articles, blogs and code, and then ran large-scale reinforcement learning.

Benchmarks and Measured Results

Official figures: FrontierCode 2.0 24.0, slightly above SWE-2's 23.8; DeepSWE 1.1 60.1; Terminal-Bench 2.1 89.2; SWE-Bench Pro v2 Hard 89.9. Versus the Inkling-Small base, Terminal-Bench 4.0 rises from 0.9 to 20.0. These are Cognition's own evaluations with no third-party replication yet.

Getting Started for Developers

Pick SWE-2 Flash in Devin Desktop or Devin CLI; it is free for Pro, Max and Teams users until the end of November. It suits code review and batches of small changes, while harder tasks can stay with SWE-2 or other frontier models. Weights are not downloadable; access will come through the secure sharing program.