On November 4, Cognition released the coding model SWE-2 Flash: three times as fast as its flagship SWE-2 and one-fifth the price, free in Devin Desktop and the command line through November 30. It targets everyday engineering work — automations, code review, feature development, documentation updates — rather than beating the flagship across every benchmark.
The speed and price are real; the scores need a discount
Flash is built on Inkling-Small from Thinking Machines, a sparse mixture-of-experts base with 276 billion total parameters and 12 billion active per token, to which Cognition added mid-training and reinforcement learning. Its own numbers show Terminal-Bench 4.0 rising from the base model's 0.9% to 20.0%, still below SWE-2's 27.3%; FrontierCode 2.0 comes in at 24.0%, slightly ahead of SWE-2's 23.8%. All of these are Cognition's own measurements, with no independent evaluation yet.
Why the weights stay private
Unlike open-weight models such as Kimi K3.1 Pro, SWE-2 Flash's weights are not being released for now. Cognition cites cybersecurity: targeted security training raised its ExploitBench score from the base model's 11.3% to 34.8%, and the same capabilities could also help attackers. The company is building a secure sharing program that opens the weights only to trusted partners and research institutions. Per-token pricing and the context window have not been announced either.
For developers, Flash's value is not its parameter count but its cost per task: if everyday jobs run at one-fifth the price and three times the speed, Devin's subscription math looks much better. The split is clear, too: the flagship SWE-2 remains the big model at 2.8 trillion total parameters and 104 billion active, reserved for the hardest jobs, while Flash handles the daily work — a pairing, not a replacement. The catch is that Cognition's self-reported scores still have to hold up on real codebases, and only user testing will prove that. The same race is consolidating fast: Meta has just acquired Conductor, a tool for parallel multi-agent coding, and competition among coding agents is shifting from single models to whole workflows.
