Cognition Releases SWE-2 Flash Coding Model, Built on Inkling-Small with New…
Cognition has released the fast coding model SWE-2 Flash, built on thinkymachines' Inkling-Small (276B total parameters, 12B active parameters), at 5x lower cost and 3x faster than SWE-2. The model adds a mid-training stage for the first time before RL, training on trillions of tokens of code, scientific articles, and blogs, saving the scarcest high-quality data for last. Cognition says it surpasses all US open-weight models on major benchmarks, including models with 3x more parameters, and plans to safely share the weights with partners and researchers through the program; Pro, Max, and Teams users of Devin Desktop and Devin CLI can use it for free until the end of this month.
Why read this
SelectedAI score 70
Cognition has published the training pipeline and base model parameters of SWE-2 Flash, allowing readers to understand the source of its cost and speed advantages.
Topics
Body
For the first time, we also added a mid-training stage before RL. We trained on several trillion tokens of code, scientific articles, blogs, and more, saving the scarcest, highest-quality data for last. https://t.co/zi2nkasAaz