[Cognition Releases SWE-1.7, a Programming Large Model with 1000 Tokens per Second] According to monitoring by Beating, Cognition has launched a new AI programming large model, SWE-1.7, which is now live on Devin. SWE-1.7 is developed based on the Kimi K2.7 Code from the Dark Side of the Moon. In multiple intelligent programming tests, it approaches the performance levels of GPT-5.5 and Claude Opus 4.8, while significantly reducing generation costs. By integrating Cerebras inference chips, SWE-1.7 achieves a processing speed of 1000 tokens per second. To handle ultra-long tasks, the model has learned to automatically summarize the current state at contextual critical points and continue running from the summary. This enables single-task durations of up to 6 hours. The team introduced an alternating length penalty mechanism during training, penalizing redundant outputs at specific stages to encourage the model to simplify reasoning for straightforward tasks while reserving extended reasoning for complex problems. To overcome the limitations of single-cluster computing power, the team also built a multi-cluster distributed training architecture spanning three continents. By synchronizing only weight changes, global node parameter updates can be completed within minutes. [Original Link]