KAT-Coder-V2.5-Dev
3B active parameters doing 69% of SWE-Bench's homework
What it is
Kuaishou's Kwaipilot team fine-tunes Qwen3.6-35B-A3B into a 35B-total / 3B-active agentic coder with 262K context, optional thinking mode, native tool calling, and reasoning traces that persist across turns — built with SFT on 127K agent trajectories plus reinforcement learning.
Why it's interesting
69.40 on SWE-Bench Verified from 3B active parameters is the best efficiency ratio we've cataloged, beating peers its size and several models well above it. It's the strongest evidence yet that agentic coding skill distills down to hardware ordinary teams actually have.
Use cases
- Local coding agents via GGUF builds in Ollama, LM Studio, and Jan
- Cheap self-hosted backends for SWE-agent-style pipelines
- RL-for-agents research on an accessible base
Who it's for
Coding-agent builders on modest hardware; local-AI tinkerers
Setup
Easy. vLLM/SGLang on one GPU for BF16; community GGUF quants run under llama.cpp, Ollama, LM Studio, and Jan
Limitations & cautions
Tuned hard for agentic coding — general chat is not the point. Launch numbers are self-reported, independent evals are early, and the missing LICENSE file should bother you exactly as much as it bothers us.
Editorial takeaway
The trillion-parameter arms race gets the headlines. The 3B-active model quietly closing SWE-Bench tickets gets the deployments.