Models

KAT-Coder-V2.5-Dev
Lightweight agentic coding.
- Open-weight 35B-A3B MoE for long-horizon coding.
- Up to 1M tokens; better tool use, less repetition.
- SWE-bench Verified: 69.4% · PinchBench: 93.4%.

KAT-Coder-V2.5
Trillion-parameter coding intelligence.
- Beyond GLM-5.1, approaching GLM-5.2.
- SWE-bench Pro: 65.2% (Opus 4.8: 69.2%).
- PinchBench: 94.2% (Opus 4.8: 93.5%).

KAT-Coder-V2
Reliable software engineering.
- #6 on the Artificial Analysis Coding Index (Mar 2026).
- SWE-bench Verified: 79.6% (Opus 4.6: 80.8%).
- PinchBench: 88.7%, ahead of GLM-5 and MiniMax M2.7.