ChainThink reports that on July 9, according to official announcements, Cognition has released the programming large model SWE-1.7, which is now live on Devin.
This model is developed based on Moonshot's Kimi K2.7 Code and performs close to GPT-5.5 and Claude Opus 4.8 across multiple agent programming benchmarks, while offering lower generation costs.
It has been disclosed that after integrating SWE-1.7 with the Cerebras inference chip, it can achieve a processing speed of 1,000 tokens per second.
To handle extremely long tasks, the model can automatically summarize the current state at context breakpoints and continue processing based on the summary, enabling single tasks to last up to 6 hours.
In addition, the team introduced an alternating length penalty mechanism during training to suppress redundant outputs at specific stages, thereby compressing reasoning for simple tasks and reserving lengthy reasoning for complex problems.
Its training architecture also employs a multi-cluster distributed solution across three continents, enabling global node parameter updates within minutes by synchronizing only weight changes.