Skip to main content
Back

DavidAU's Qwen3.8 Fine-Tune Beats the Base Model at a Tenth of Thinking Tokens

A community fine-tune of Qwen3.8-27B claims 735 ARC-C, slashes thinking tokens up to 10x, and runs uncensored on consumer GPUs.

  • DavidAU released a multi-stage fine-tune of Qwen3.8-27B claiming 735 ARC-C and 882 ARC-E scores.
  • TURBO branding reflects thinking token reductions of one-half to one-tenth versus base Qwen3.8.
  • Built with COLD FUSION (custom GAIN method plus Unsloth) and Fable Fusion 711 training techniques.
  • De-censored via Heretic with Arbitrary-Rank Ablation, dropping refusals from 99/100 to 11/100.
  • 4-bit quant reportedly hits 99% of 8-bit performance; MTP variants exceed 90 t/s on a 5090.
  • Apache 2.0, 256k context, vision-capable, runs in llama.cpp, Ollama, LM Studio, and vLLM.

A community fine-tune of Qwen3.8-27B is climbing the Hugging Face trending charts with aggressive claims: higher benchmark scores than the base model, a fraction of the reasoning tokens, and a footprint that fits on a single consumer GPU. Qwen3.8-27B-TURBO-Fable-Cold-Fusion from DavidAU is a multi-stage tuned, de-censored, GGUF-packaged variant built outside any major lab.

What the benchmarks actually say

The headline pitch is a jump on ARC-Challenge and ARC-Easy. At 8-bit (mxfp8), the model reports 0.735 ARC-C and 0.882 ARC-E, versus the base Qwen3.8-27B at 0.591 and 0.782. At 4-bit (mxfp4) those numbers land at 0.719 ARC-C and 0.887 ARC-E, close enough to 8-bit performance that the author frames 4-bit as roughly 99% of 8-bit quality. For anyone running locally on 16-24 GB of VRAM, that gap matters. The benchmarks were run by the author's collaborator Nightmedia using standard multiple-choice suites (ARC, BoolQ, HellaSwag, OBQA, PIQA, WinoGrande), not a third party, and cover classification-style tasks rather than modern agentic or coding evaluations.

Slashing the thinking tokens

Qwen3.8 defaults to heavy chain-of-thought reasoning, and this fine-tune directly targets that overhead. The TURBO designation covers reductions in thinking tokens ranging from one-half to as low as one-tenth of the original, across all three reasoning modes (xhigh, medium, low), while claiming to preserve output detail and quality. Multi-turn refinement stages reportedly see the steepest cuts, sometimes reaching one-fifth the size of typical Qwen thinking blocks.