KwaiKAT Team Releases KAT-Coder-V2.5: An Agentic Coding Model Trained on 100,000+ Verifiable Repository Environments

TL;DR AI
2 min readKey summary
Kuaishou’s KwaiKAT team released KAT-Coder-V2.5, an agentic coding model built to operate inside executable repositories.
The team also open-sourced KAT-Coder-V2.5-Dev on Hugging Face under Apache-2.0.
Training data was grounded in more than 100,000 verified environments mined from real pull requests and commits.
The pipeline used automated environment construction, test-based verification, and process filtering to improve data quality.
The release points to a more reliable way to train and evaluate coding models for real software engineering tasks.

