Kwaipilot's agentic-coding generation trained on 100K+ AutoBuilder-generated verifiable repository environments across 12 languages: 127K-example SFT plus asymmetric PPO with a privileged-context critic. Proprietary Pro/Air tiers shipped via API July 10; the open-weight V2.5-Dev followed July 26.

Outputs 3

KAT-Coder-Pro / Air V2.5

model

Proprietary API tiers. SWE-Bench Pro 65.2 (second only to Claude Opus 4.8's 69.2 at release), Terminal-Bench 2.1 60.7. Base undisclosed; the KAT line post-trains external bases.

Architecture MOE

KAT-Coder-V2.5-Dev

model

Open-weight sibling, Apache 2.0, text-only: "we adopt the widely recognized Qwen3.6-35B-A3B as the base model for post-training." 35B total / 3B active, 262K context. SWE-bench Verified 69.4, SWE-bench Pro 45.96. 15K downloads in its first ten days.

Architecture MOE
Context window 262,144
License Apache 2.0
Base model qwen3.6-open

KAT-Coder V2.5 Technical Report

paper
codingagentsmoeopen-weight

Related