Skip to main content
Back

Qwen3.5-27B Distilled With Claude Opus Reasoning Hits 90K Downloads

A community fine-tune ports Claude 4.6 Opus reasoning traces into Qwen3.5-27B, delivering stable tool calling and 9+ minute autonomous agent runs on a single RTX 3090.

  • Qwen3.5-27B fine-tuned on Claude 4.6 Opus reasoning traces, released as GGUF under Apache-2.0.
  • Runs at 16.5 GB VRAM in Q4_K_M with full 262K context on a single RTX 3090.
  • Generation speed of 29 to 35 tok/s with structured think-tag reasoning.
  • Fixes Jinja template crash on the developer role used by Claude Code and OpenCode.
  • Sustains 9+ minutes of autonomous agent execution without stalls; only 27B variant shows stable tool calling.
  • Trained with Unsloth LoRA SFT on Opus-4.6-Reasoning-3000x-filtered; full guide on GitHub.

A community developer has released Qwen3.5-27B-Claude-4.6-Opus-Reasoning-Distilled, a GGUF-packaged fine-tune that transplants Claude 4.6 Opus reasoning patterns onto the Qwen3.5-27B dense base. The pitch is simple: keep the local, quantized, run-it-on-one-GPU footprint of Qwen, but inherit the structured thinking style that makes Opus feel competent inside coding agents. The release has already crossed 90,000 downloads on the Hub.

What's actually inside the release

The model is a supervised fine-tune of Qwen3.5-27B trained on two distillation datasets: nohurry/Opus-4.6-Reasoning-3000x-filtered, which provides Claude 4.6 Opus reasoning trajectories, and Jackrong/Qwen3.5-reasoning-700x for additional structured step-by-step samples. It ships under Apache-2.0 with three quantizations available directly from Hugging Face:

  • Q3_K_M at 13.3 GB
  • Q4_K_M at 16.5 GB
  • Q8_0 at 28.6 GB

Getting it running takes one command in any major local runtime. For llama.cpp:

llama serve -hf Jackrong/Qwen3.5-27B-Claude-4.6-Opus-Reasoning-Distilled-GGUF:Q4_K_M

Ollama, LM Studio, Jan, vLLM, SGLang, and Docker Model Runner all have first-class entries too, so you can point Claude Code, OpenCode, or any OpenAI-compatible agent at