AI Research Atlas

KAT-Coder-V2.5 and V2.5-Dev

Kuaishou (Kwaipilot) · 23 July 2026

Kuaishou's proprietary agentic coder KAT-Coder-V2.5 (July) got a 35B-total, 3B-active open sibling, V2.5-Dev, post-trained from Qwen3.6-35B-A3B.

The open Dev model reports 69.4 on SWE-bench Verified against Qwen3.5-27B's 68.6 in its own table and cuts abnormal tool-label outputs from 9.34% to 0.28%. Apache-2.0, text-only.

Date
Thursday, 23 July 2026
Lab
Kuaishou (Kwaipilot)
Kind
open-weights
Access
open weights

Figures

MeasureValueMeasured by
SWE-bench Verified (Dev)69.4
vs Qwen3.6-35B-A3B 64.4
company

Open weights dated by HF repo (2026-07-23); the proprietary V2.5 appeared on OpenRouter 2026-07-10. Benchmarks are Kuaishou's.

Sources

  1. huggingface.co/Kwaipilot/KAT-Coder-V2.5-Dev
  2. openrouter.ai/api/v1/models

This record was checked against its sources on 6 October 2026. How we check

Related