Core insight
Most Qwen3.8‑27B derivatives focus on raw bench gains; this build instead targets reasoning efficiency and deployability. By combining COLD FUSION (GAIN+Unsloth) and Fable‑Fusion stages, the authors reduce per‑turn reasoning/thinking token blocks (typical reductions to 1/5 — sometimes 1/10 of vanilla sizes) while preserving or improving downstream benchmark scores — enabling faster, lower‑cost multi‑turn and MTP inference on limited hardware.
Key capabilities
- Benchmarks: reported 8‑bit ARC‑C ≈735 and ARC‑E ≈882; 4‑bit ARC‑C ≈719, exceeding base Qwen3.8‑27B in the listed metrics. 4‑bit performance reported near 99% of 8‑bit in tests.
- Thinking control: supports Qwen’s xhigh/medium/low reasoning modes with modified Jinja templates to tune preserved thinking and injected system prompts.
- Quant formats: multiple GGUF releases (regular and MTP) using Neo‑Imatrix quants; some quants use mixed precision (output tensor partly 16‑bit) for accuracy vs memory tradeoffs.
- Multi‑token prediction (MTP): MTP quants aim for higher throughput (reported t/s gains) when token acceptance rates are adequate; guidance provided on temp/rep‑pen tuning.
- Vision & context: multimodal support (image/video) with separate mmproj file; native large context support (256k, extensible via YaRN up to 1M tokens with overrides).
- Accessibility: packaged for local consumer GPU use (GGUF), aimed at LMStudio, text‑generation‑webui, oobabooga, etc.
What sets it apart
Rather than blind bench‑maxing, this fine‑tune explicitly rearchitects the model’s reasoning blocks to: (1) shorten internal deliberation tokens, (2) maintain/output high detail, and (3) offer both decensored (Heretic) and tuned variants. It bundles both regular and MTP Neo‑Imatrix quants so users can pick speed vs acceptance tradeoffs on their hardware.
Who it’s for and trade‑offs
Great fit if you need: local multimodal inference with improved throughput on consumer GPUs; fast multi‑turn agent experiments; creative or coding workflows where shorter internal thinking reduces latency. Look elsewhere if you require official, supported production weights from the Qwen team: this is a community multi‑stage fine‑tune and includes a strong “heretic/uncensored” variant, which reduces refusal rates and can generate explicit content—test carefully for safety, compliance and alignment with your deployment policy.
Practical notes & limitations
- Decensoring: Heretic steps lower refusal rates (KLD reported improvements) but change refusal behavior—evaluate for safety and legal constraints.
- App settings: MTP quants and thinking modes may need UI/harness changes (chat_template.jinja, token‑config.json) and specific sampling settings for best results.
- Vision: requires the separate mmproj artifact to enable image processing.
- Not an official Qwen release: community fine‑tune maintained by DavidAU and collaborators; license on Hugging Face lists Apache‑2.0 but verify for your use case.
Where it fits
Positioned between base Qwen3.8‑27B and experimental bench‑maxed forks: it prioritizes usable speed/efficiency improvements for local deployment while still improving core benchmarks, making it a pragmatic option for local multimodal and agentic experiments that need lower latency and MTP throughput.
