SWE-2
SWE-2 builds on the 2.8-trillion-parameter Kimi K3 model. Its reinforcement learning trains all reasoning-effort levels in one run.
What the release provides
Training uses a different cost penalty for each effort level. Medium, High and Max adjust the balance between task exploration and execution. The release also adds training environments and improves the checks used to judge completed tasks.Cognition: SWE-2 training and release
How to use it well
Practical guidance. Choose effort from the complexity of the requested change. Inspect both the result and the tool actions; longer reasoning alone does not show that the software works.
Specifications & access
| Detail | Value and conditions |
|---|---|
| Access | Provider product |
| Current access | Available through Devin products, with Medium, High and Max settingsDevin: Current model catalog |
Related models
Kimi K2.7-based coding model trained to summarize its working context and continue long software-engineering tasks.
Launch date not confirmedProvider product- SWE-2 Flash
Cognition · SWE
A coding model based on Thinking Machines’ Inkling-Small.
Launch date not confirmedProvider product