Cognition releases SWE-2 coding model
Cognition AI, Inc. →Cognition introduced SWE-2, a coding model trained using a new reinforcement-learning approach across multiple reasoning-effort levels. The company said the model was post-trained from Kimi K3 and extended its training infrastructure into the multi-trillion-parameter range. The release described changes to cost penalties, reward baselines, rollout serving and training data, including a larger set of reinforcement-learning environments. SWE-2 became available in Devin Desktop and CLI, with rollout to Devin Web and Fusion also announced. Cognition presented benchmark results and pricing comparisons as evidence of improved coding capability and efficiency. These figures are company-reported evaluations, rather than an independently established guarantee of customer performance or savings.