KAT-Coder-V2.5-Dev brings 3B active-parameter agentic coding to 256K context
airesearch12 · x · 2026-07-28
KAT-Coder-V2.5-Dev targets agentic coding with 3B active parameters
KwaiAICoder’s new KAT-Coder-V2.5-Dev is a 35B open-weight MoE coding model built on Qwen3.6-35B-A3B. It uses only 3B active parameters at inference time but is positioned as a same-scale leader for agentic coding.
Reported results
- SWE-bench Verified: 69.4
- SWE-bench Multilingual: 63.0
- Terminal-Bench 2.1: 41.02
- SWE-bench Pro: 46.0
- PinchBench: 93.4
- SciCode: 44.2
- KAT Code Bench: 46.2
Technical details
- Native 256K context
- Tool calling
- Thinking / non-thinking modes
- Optional reasoning-history preservation
- Trained with 127K SFT samples + RL
- Abnormal tool labels dropped from 9.34% to 0.28%
- Continuous repetition reduced to 0%
- Open-weight, text-only, Apache 2.0
More from coding & agent
- AJ Vaynerchuk uses Claude Code to turn a 45-second memo into internal apps — tomcrawshaw01 · 2026-07-28
- Reverse Engineering with Claude: Using Ghidra and Frida to Rewrite C Code — yacineMTB · 2026-07-28
- Agensis turns AI agents into coworkers in a shared Slack-like workspace — jasonkneen · 2026-07-28
- Nebius launches a local relay that routes four coding agents to open models — HowDevelop · 2026-07-28
- Netlify says AI is making large code migrations more practical and less permanent — thisiskp_ · 2026-07-28
- Cursor details an agent swarm that kept context small and hit 80% on a SQLite-from-docs benchmark — bibryam · 2026-07-28