Cognition Releases SWE-2 Flash: Mid-Training on Thinking Machines' Inkling-Small, One-Fifth the Price of SWE-2, Weights Held Back Over Cyber Risk

A smaller SWE-2 for everyday engineering work

Cognition, the company behind Devin, announced SWE-2 Flash on its official blog on November 4. The positioning is straightforward: much of what software agents do day to day, such as automations, code review, feature development and documentation updates, can increasingly be handled by smaller models, while harder jobs stay with the much larger SWE-2. SWE-2 Flash went live the same day in Devin Desktop and CLI. Cognition puts it at one-fifth the price of SWE-2 and about three times the speed; the blog's demo replays the same Terminal-Bench 4.0 task at a simulated 200 tokens per second versus SWE-2's 60. The page does not list per-token pricing, and the model is free to use through November 30.

A new base model, and mid-training for the first time

SWE-2 Flash is built on Thinking Machines' Inkling-Small, with 276B total parameters and 12B active per token; for comparison, SWE-2 has 2.8T total and 104B active. Cognition calls it "a fully American coding model." The change in the recipe is that Cognition added a large-scale mid-training stage before reinforcement learning for the first time. It first compared the token efficiency of candidate data mixes against a held-out set of real-world software engineering data, then settled on a few trillion tokens of high-quality scientific articles, blogs, code and other sources, trained in three stages with the scarcest, highest-quality data saved for later. The reinforcement learning recipe changed only slightly; most of the gains came from the quality and diversity of the RL tasks.

According to Cognition, mid-training plus RL raised Terminal-Bench 4.0 from the base model's 0.9% to 20.0% and FrontierCode 2.0 from 3.2% to 24.0%.

The other columns behind "beats American open models"

Cognition says SWE-2 Flash outperforms all American open-weight models on leading evals, even models three times larger (Muse Spark Open in the table has 816B total parameters). The same comparison table lists other models too: on Terminal-Bench 4.0, GLM 5.3 Flash scores 41.8, DeepSeek V4.1 Flash 1017 scores 42.3 and Claude Haiku 5.5 scores 39.2, all well above SWE-2 Flash's 20.0; on FrontierCode 2.0, Claude Haiku 5.5's 27.7 is also higher. On SWE-Bench Pro v2 Hard, SWE-2 Flash scores 89.9, close to SWE-2's 90.6. All of these scores were published by Cognition and have not been independently reproduced.

Why the weights are not out

Cognition specifically targeted cybersecurity work in training, raising ExploitBench from the base model's 11.3% to 34.8%. It says the same capabilities can be used by attackers, so it is not sharing the weights publicly yet; instead it is building a program to share them securely with trusted industry partners and research labs, with details promised "in a few weeks." For developers, SWE-2 Flash is for now only usable inside Devin's products. The more interesting signals are that a leading coding-agent company is moving everyday traffic onto a small model, and that "mid-train first, then RL" took a base model with very low starting scores to a usable level.