Models
KAT-Coder-V2.5-Dev
Lightweight agentic coding.
- Lightweight 35B-A3B MoE with open weights, built for long-horizon agentic coding.
- Native support for up to 1M tokens, with more reliable tool use and fewer repetitive behaviors.
- 69.4% on SWE-bench Verified and 93.4% on PinchBench.
KAT-Coder-V2.5
Trillion-parameter coding intelligence.
- Trillion-parameter coding model, surpassing GLM-5.1 and approaching GLM-5.2.
- 65.2% on SWE-bench Pro, compared with Claude Opus 4.8 at 69.2%.
- 94.2% on PinchBench, compared with Claude Opus 4.8 at 93.5%.
KAT-Coder-V2
Reliable software engineering.
- Ranked 6th on the Artificial Analysis Coding Index in March 2026.
- 79.6% on SWE-bench Verified, close to Claude Opus 4.6 at 80.8%.
- 88.7% on PinchBench, surpassing GLM-5 and MiniMax M2.7.