Cognition has launched SWE-2 Flash, a new coding model that increases token generation speed threefold and reduces operational costs to one-fifth the price of the company’s standard SWE-2 tool. Built on the Inkling-Small architecture from Thinky Machines, the new model utilizes 12B active parameters and is available for free to users on the Pro, Max, and Teams plans of the Devin coding agent through Nov 30.
The model was developed over six weeks using reinforcement learning and a first-time “mid-training” stage to refine the base model. While Cognition claims the tool surpasses all American open-weight models of comparable size, some critics allege that the firm’s reported benchmarks are inconsistent and omit results from other open models that outperform the new release. To mitigate the cybersecurity risks of open weights, the company is creating a program to share access with specific industry partners and researchers.
