GitHub radar
Qwen3.8-27B Cold Fusion: a more speedy brain, works offline
DavidAU finetuned Alibaba’s Qwen3.8-27B to reason 50%–90% fewer steps. GGUF formatted. Runs locally on LM Studio or Ollama.
Qwen3.8-27B is a strong multimodal model developed by Alibaba, but it tends to take many steps to reason through each response. This makes it great for analytic situations, but a bit too much for general conversation. DavidAU finetuned this model with Cold Fusion + GAIN to reduce its reasoning steps by 50%–90% without sacrificing accuracy. MTP variations can be used to further boost the token rate to over 90 tokens/sec. It is distributed as GGUF and runs locally on LM Studio, Ollama, and llama.cpp. A context size of 256,000 tokens (up to 1 million). You can adjust the number of steps taken during each request based on your desired level of speed vs. accuracy.
Why a vibe-coder should care
Same Qwen3.8-27B that you might be familiar with, but faster at general questions. Runs completely offline. No subscription fees or API keys needed. With more than 238,000 downloads in 30 days, we’ve found the community wants to run their own version of Qwen3.8 locally.
How to install
Copy this and send it to your agent — Claude Code, Codex, any of them:
Set up Qwen3.8-27B Cold Fusion by DavidAU locally: https://huggingface.co/DavidAU/Qwen3.8-27B-Cold-Fusion-GAIN-V1.1-NM-DAU-NEO-MAX-MTP-GGUF — pick the right GGUF quantization for my machine, run it via Ollama, and show me how to switch between reasoning modes
A Mac with 32 GB of RAM or more, or a GPU with 16 GB of VRAM — in 4-bit version the model takes up about 16 GB.
Open on Hugging Face▌ More finds