GitHub radar
Qwen3.8-27B retuned to think less — and run faster
ukisai fine-tuned Alibaba's Qwen3.8-27B to stop over-reasoning: 58% fewer thinking tokens and faster inference — available as a local GGUF.
ukisai fine-tuned Alibaba's Qwen3.8-27B — a multimodal 27B model that handles text and images — to reduce the verbose reasoning chains that slow it down. The approach: penalize the marker tokens that trigger long thinking rollouts, and borrow transfer learning from BottleCap AI's ThinkingCap. The result: 58.3% fewer thinking tokens (median), noticeably faster inference, and a LiveCodeBench score that actually improved from 76.76% to 81.55%. Available as a GGUF for llama.cpp, Ollama, and vLLM. License is Swift Open License 1.0 — free for individuals and businesses under $1M in annual revenue; commercial users above that threshold need separate terms.
Why a vibe-coder should care
If you already run Qwen3.8-27B locally, this variant gives meaningfully faster responses without changing hardware. For local agent tasks and repeated text work the speed-up is real. Honest caveat: it's still a local 27B model — quality is what it is, not a substitute for frontier models.
How to install
Copy this and send it to your agent — Claude Code, Codex, any of them:
Set up the Swift-Qwen3.8-27B model locally via Ollama: https://huggingface.co/ukisai/Swift-Qwen3.8-27B-GGUF — pick the right quantized variant for my machine, install it, and show me how to use it.
You need a Mac with 32 GB of memory or a GPU with 24 GB of VRAM — the 4-bit version takes about 16 GB.
Open on Hugging Face▌ More finds