Image: Cognition
Cognition, the company behind the Devin coding agent, has released SWE-2 Flash, a smaller and cheaper version of its SWE-2 coding model meant for routine engineering work. Cognition says it runs at three times the speed of SWE-2 for one fifth of the price, and that it beats every American open weight model on the coding tests the company ran.
SWE-2 Flash is not a model built from scratch. According to Cognition's announcement, it starts from Inkling-Small, an open weight model from Thinking Machines with 276 billion total parameters, of which 12 billion are active for any given token. Cognition added a mid-training stage on a few trillion tokens of scientific articles, blogs, code and other sources, then ran large scale reinforcement learning on coding tasks. The company says this is the first time its training recipe has included a mid-training step.
Introducing SWE-2 Flash, our new fast & efficient coding model.
— Cognition (@cognition) November 4, 2026
On leading evals, it outperforms all American open-weight models, even ones 3x larger.
To balance transparency and cyber risks, we are building a program to securely share the weights with partners & researchers. pic.twitter.com/MJQ2c1HKIi
The model is aimed at the everyday jobs that coding agents handle, which Cognition lists as automations, code review, feature development and documentation updates. For harder problems, Devin users can still pick the full SWE-2, which Cognition lists at 2.8 trillion total and 104 billion active parameters. The SWE family also includes the older SWE-1.6 and SWE-1.7 models, plus SWE-1.7 Lightning, a faster version served on Cerebras hardware, according to Devin's model documentation.
Access is limited to Cognition's own tools for now. SWE-2 Flash is available in Devin Desktop and the Devin CLI, where users choose it from the model selector. RuntimeWire reports that it is free for paid Devin users through November 30. Cognition has not published a per token price for Flash, and there is no public API endpoint or model ID outside Devin at the time of this writing. For comparison, Devin's documentation lists SWE-2 at $3 per million input tokens and $15 per million output tokens, so the one fifth claim would put Flash well below that if it holds once the free period ends.
Image: Cognition
Cognition's benchmark table tells a more mixed story than the headline. On FrontierCode 2.0, SWE-2 Flash scored 24.0%, slightly ahead of SWE-2 at 23.8% and OpenAI's GPT-6 Luna at 21.9%, but behind Anthropic's Claude Haiku 5.5 at 27.7%. On Terminal-Bench 4.0 it scored 20.0%, far above the 0.9% of its Inkling-Small base, yet well short of Claude Haiku 5.5 at 39.2% and SWE-2 at 27.3%.
The word American in the claim also matters. Cognition's own table shows two Chinese open weight models, GLM 5.3 Flash and DeepSeek V4.1 Flash, ahead of SWE-2 Flash on several tests. DeepSeek V4.1 Flash scored 42.3% on Terminal-Bench 4.0 and 77.0% on DeepSWE 1.1, compared with 20.0% and 60.1% for Cognition's model. American open weight rivals in the comparison, such as Beam 1.0 and Muse Spark Open, trail SWE-2 Flash on most rows despite being larger.
All of these figures are self reported by Cognition and have not been independently reproduced. The speed comparison is also an illustration rather than a live measurement. Cognition's blog shows SWE-2 Flash at 200 tokens per second against 60 for SWE-2, using transcripts from a single Terminal-Bench task replayed at those simulated rates.
The weights are not public. Cognition says it trained the model for cybersecurity work, raising its ExploitBench score from 11.3% to 34.8%, and that because attackers could use the same abilities, it is building a program to share the weights securely with trusted industry partners and research labs. The company says more details will come in a few weeks.
For Devin users, the practical takeaway is that SWE-2 Flash is worth trying on routine tasks while it is free, with SWE-2 or other models still available for harder work. Teams planning to depend on it may want to wait for Cognition to publish its price and for independent testing, such as the Artificial Analysis rankings, to show how it performs outside the company's own evaluations.
Comments