Cognition to Release SWE-1.7, a Large-Scale Programmatic Model, Emitting 1000 Tokens Per Second

According to Dynamic Insight monitoring, Cognition has released a new AI programming large model, SWE-1.7, which is now live in Devin. SWE-1.7 is developed based on the dark side of the moon's Kimi K2.7 Code. In multiple intelligent agent programming tests, it has approached the level of GPT-5.5 and Claude Opus 4.8 but with significantly reduced generation costs. By integrating with the Cerebras inference chip, SWE-1.7 has achieved a processing speed of 1000 Tokens per second.

To handle ultra-long tasks, the model has learned to automatically summarize the current state at context critical points and continue running from the summary. This allows the duration of a single task to reach 6 hours. The team introduced an alternating length penalty mechanism during training, penalizing redundant outputs at specific stages, prompting the model to streamline inference for simple tasks and reserve long inference for challenging ones. To overcome a single cluster's computational power limit, the team has also built a multi-cluster distributed training architecture spanning three continents, achieving global node parameter updates in minutes by synchronizing only the weight change amounts.