SWE-2 Flash coding model released by OpenAI

Cognition has introduced SWE-2 Flash, a new coding model featuring a large-scale mid-training stage built on top of Inkling-Small, followed by reinforcement learning. The company states that the model achieves high speed and low cost while outperforming larger American open-weight models on leading benchmarks. To balance transparent access with cybersecurity risks, Cognition is building a secure program to share the model weights with trusted industry partners and researchers rather than releasing them publicly. Open article →