The developer of the SWE-2 Flash AI coding tool will provide free access to the software for the next month, promising performance that rivals top tier closed-source models such as Luna and Haiku. Cognition said the efficiency-focused model was created by using a technique known as “mid-training” to build upon the Inkling-Small architecture.
Designed to be faster and cheaper than the standard SWE-2, the new tool reportedly outperforms all American open-weight coding models on industry benchmarks, including rivals that are 3x larger in size. The firm is currently establishing a weight-sharing program for selected researchers and partners to maintain transparency while mitigating cybersecurity risks.
