On November 4, 2026, Cognition — the startup behind the AI software engineer Devin — released SWE-2 Flash, a coding model built to run three times faster and cost one-fifth as much as its own flagship, SWE-2, while beating every other American open-weight model on the benchmarks the company tracks. It shipped free, inside Devin Desktop and Devin CLI, through November 30.
A smaller model doing a bigger model’s job
SWE-2 Flash isn’t a scaled-down version of SWE-2 in the usual sense — it’s built on a different, much smaller foundation. Where SWE-2 runs on a 2.8-trillion-parameter model with 104 billion active at a time, Flash is built on Inkling-Small, a model from Thinking Machines (the AI lab co-founded by former OpenAI CTO Mira Murati) with 276 billion total parameters and only 12 billion active. In Cognition’s own demo, replaying the same Terminal-Bench 4.0 task, SWE-2 Flash generates at roughly 200 tokens per second against SWE-2’s 60.
The pitch, in Cognition’s own words, is that “smaller models are increasingly able to handle a significant portion of the work that software agents do” — automations, code review, routine feature development, documentation updates — the high-volume, lower-stakes end of the job that doesn’t need a frontier-sized model to get right. The economics matter here as much as the capability: a model that’s a fifth the price turns a lot of previously uneconomical agent workflows (running an agent continuously in the background, re-checking every pull request) into something a team can actually afford to leave running.
What’s actually new in the training recipe
Cognition says the gains didn’t come from reinforcement learning alone — the part of training where a model is rewarded for successfully fixing bugs or passing tests. Before that stage, Flash went through what the company calls a new “mid-training” stage: a few trillion tokens of scientific articles, blogs, code, and other data, fed in three passes with the scarcest, highest-quality data saved for last. The reinforcement-learning stage after that reused most of SWE-2’s own recipe, with the improvement credited mainly to better, more diverse training tasks — audited for correct instructions, realistic environments, and reliable verifiers, rather than just more volume.
The jump from the raw Inkling-Small base model to the finished SWE-2 Flash is the clearest evidence the recipe worked: on Terminal-Bench 4.0, the base model scored 0.9%; Flash scores 20.0%. On SWE-Bench Pro v2 Hard, a benchmark of realistic, PR-sized coding tasks, the base model managed 56.5; Flash hit 89.9. On FrontierCode 2.0, Flash’s 24.0 edges out even SWE-2’s own 23.8 — a rare case of the cheaper, faster model narrowly beating its bigger sibling on a benchmark, even though SWE-2 still leads comfortably on Terminal-Bench 4.0 (27.3) and DeepSWE 1.1 (73.0 vs. Flash’s 60.1).
The part that’s deliberately not public: the weights
Cognition is explicitly not releasing SWE-2 Flash’s weights to the public, and says so plainly: it’s citing cyber-attack risk as the reason. That’s not an abstract concern for this particular model — training it for stronger coding ability also measurably improved its exploit-development skill, with ExploitBench scores climbing from 11.3% at the base model to 34.8% after training. Instead of an open release, Cognition says it’s building a program to share weights “securely with trusted industry partners and research labs,” with enterprises and governments getting on-prem control and audit visibility, and more detail promised “in a few weeks.” No date has been set.
Why this is worth watching past the free trial
The free window through November 30 is a trial, not the business model — Cognition hasn’t announced standalone API pricing or a public release date for access outside Devin. But the move fits a pattern that’s been building across the industry all year: labs splitting their lineups into an expensive, maximally capable flagship and a cheap, fast sibling built specifically for the repetitive, high-volume slice of agentic work — the same logic behind OpenAI’s and Anthropic’s own “mini” and “flash” tiers, just applied here to a coding-specific agent rather than a general chatbot. If Cognition’s benchmark claims hold up under independent testing once the free period ends and real usage data comes in, SWE-2 Flash would mark one of the more concrete signs yet that a meaningfully smaller model, trained the right way, can now do most of a software agent’s day-to-day job for a fraction of the cost — without needing the giant model running underneath it at all.
Image: Lukas Blazek via Pexels.
