CognitionSWE
SWE-1.7
SWE-1.7 adds reinforcement learning to Kimi K2.7. It learns to summarize its working context and continue long engineering tasks.
What the release provides
The training targets reliable coding work across many actions. Learned context compaction keeps a task moving after its earlier working context is condensed. The current model catalog retains it as the previous flagship.Cognition: SWE-1.7
How to use it well
Practical guidance. Review the saved task state after a long run. A compacted summary can omit requirements, so compare the final change with the original request.
Specifications & access
| Detail | Value and conditions |
|---|---|
| Access | Provider product |
| Current listing | Listed as a previous flagship in the current Devin model catalogDevin: Current model catalog |
Related models
A software-engineering model with standard and Fast serving options.
Launch date not confirmedProvider productA coding-agent model post-trained from Moonshot’s Kimi K3.
Launch date not confirmedProvider product