88Z AI, 10 July — Kuaishou KwaiKAT has released its flagship Agentic Coding model, KAT-Coder-Pro V2.5, achieving breakthroughs in long-horizon engineering, general agentic capabilities, and large-scale reinforcement learning. By constructing over 100,000 runnable repositories via AutoBuilder and employing multi-framework RL with Mixture of Experts techniques, the model leads in benchmarks such as SWE-Bench Pro.
At a glance
- KAT-Coder-Pro V2.5 achieves breakthroughs in long-horizon engineering and general agentic capabilities
- Leads in multiple benchmarks based on AutoBuilder and Mixture of Experts techniques
The update
The KwaiKAT team at Kuaishou has officially launched KAT-Coder-Pro V2.5, a flagship Agentic Coding model. This version achieves significant breakthroughs in three key dimensions: long-horizon engineering capabilities, general agentic abilities, and large-scale reinforcement learning. To enhance performance, the team utilized the AutoBuilder automated pipeline to construct over 100,000 runnable code repository environments, providing a solid foundation for training. In terms of training strategy, KAT-Coder-Pro V2.5 employs multi-framework reinforcement learning and Mixture of Experts techniques. Thanks to these technological innovations, the model has achieved leading results in authoritative benchmarks such as SWE-Bench Pro and PinchBench, demonstrating its robust strength in complex coding tasks.
Sources & editorial note
This article is an AI-assisted synthesis by 88Z AI based on the public sources below. Product claims and availability are attributed to their publishers. It does not represent independent testing or first-hand reporting.
- KwaiKATSource published
Source dates above refer to the original announcements; the date under the headline is this article’s publication on 88Z AI.