NovaeonStudio/Qwen3.8-27B-TURBO-Fable-Heretic-oQ8-fp16-mtp

[image]

Qwen3.8-27B-TURBO-Fable-Heretic · oQ8-fp16-mtp — a Novaeon.Studio build

An uncensored, vision-capable, hard-reasoning 27B dense for Apple Silicon — with an MTP head that genuinely works under 8-bit. This is an oMLX oQ8 (near-uniform 8-bit, sensitive tensors fp16) quantization of `DavidAU/Qwen3.8-27B-TURBO-Fable-Cold-Fusion-735-882-Heretic-Uncensored-NM-DAU` — a multi-stage Cold-Fusion + Fable-Fusion merge that was subsequently de-censored with [Heretic](https://github.com/p-e-w/heretic) ARA — repacked for the Mac in oMLX-native format. Built, tuned, and benchmarked on an Apple M5 Max (128 GB) by [Novaeon.Studio](https://novaeon.studio).

Why this build exists: we wanted a local model that is (a) genuinely uncensored, (b) still a strong reasoner rather than a lobotomized one, (c) multimodal, and (d) fast enough to be usable as a dense 27B on a Mac. The native MTP head is the lever for (d) — and unlike most 8-bit repacks, this one's MTP survived quantization and more than doubles decode. Everything in the "measured" section below is from our hardware; everything else is attributed.

Best for & why it's here

Best for: genuinely uncensored creative writing, roleplay, and red-team / security research where refusal-free output matters more than raw speed — plus vision tasks. A "says-what-you-ask" companion, not a throughput workhorse.

Why we published it: a truly uncensored (Heretic-abliterated) 27B with an intact vision tower and a working MTP head (+~115% decode). It's dense, so slower than our A3B seat (~26 tok/s) — we ship it as the selectable no-refusals + vision model rather than the default agent.

Highlights

• —Uncensored, not degraded. Heretic ARA removes refusal directions without a destructive retrain. The base card reports 0/100 refusals on its stage-1 evaluation at KL divergence 0.0535; our own independent refusal probes on these quantized weights returned zero refusals.
• —Working MTP → +115% decode. The native multi-token-prediction head is not an inert graft: under oQ8 it sustains real draft acceptance. 26.2 tok/s with draft-6 vs 12.2 tok/s with MTP off.
• —Top-of-class reasoning for its size. The base is the first model of its size/type to breach ARC-C 0.735 in 8-bit (and 0.719 in 4-bit) — well clear of the Qwen3.5/3.6/3.8 27B-class baselines.
• —"TURBO" thinking budget. The base's TURBO treatment cuts reasoning-token spend to roughly 1/2 – 1/10 of a standard thinking model, which is what makes a 26 tok/s dense 27B practical for interactive use.
• —Vision intact. Multimodal (text + image → text); the vision tower is preserved from the base.

Measured on our hardware (Apple M5 Max 128 GB · oMLX)

MTP speculative decoding

[image]

_{Standalone run (our serving seat unloaded), 220-token generation, thinking off. +115% over MTP-off — i.e. the head is doing real speculative work under 8-bit, not idling.}

Scope note (honest): these throughput numbers and the refusal probes are ours. We did not run full IFEval / agentic suites against this build — so we don't quote any. The accuracy numbers below belong to the base model and are attributed as such.

Base-model benchmarks

[image]

In-class ARC-Challenge comparison (mxfp8), benchmarks by Nightmedia, measured on the base model:

Full seven-benchmark sweep on the base:

_{Reproduced from the base model card / Nightmedia benchmark runs. These were measured on the base weights, not on this oQ8 repack.}

Other base-card highlights (credit DavidAU): multi-stage Cold-Fusion + Fable-Fusion-711 fine-tune/merge; three reasoning modes (xhigh — the default — plus medium and low); "TURBO" reduced-thinking-token treatment.

Optimal oMLX settings (figured out empirically)

We swept the tuning levers on this build. Because the native MTP head genuinely accelerates decode under oQ8, the optimal profile is MTP-on with ANE-prefill off — with a working MTP, ANE-prefill's dispatch overhead is a net loss.

<details><summary>Drop-in <code>~/.omlx/model_settings.json</code> entry</summary>

{
  "max_context_window": 262144,
  "mtp_enabled": true,
  "mtp_num_draft_tokens": 6,
  "qwen35_ane_prefill_enabled": false,
  "turboquant_kv_enabled": true,
  "turboquant_kv_bits": 8.0,
  "turboquant_skip_last": true,
  "qwen35_oq_a8_enabled": false,
  "dflash_enabled": false
}

</details>

How it was built: oMLX admin oQ conversion with oq_level: 8, dtype: float16, and — critically — `preserve_mtp: true`. Without that flag the mtp.* tensors are dropped and the speculative head is gone; with it, all 29 MTP tensors ride along and oMLX reports mtp_compatible: true.

On the architecture: qwen3_5 dense hybrid — most layers use GatedDeltaNet linear attention with periodic full-attention layers interleaved. Only the full-attention layers carry a growing KV cache, so memory stays modest even far into the 256k window.

Quickstart

# oMLX (recommended, Apple Silicon)
omlx serve NovaeonStudio/Qwen3.8-27B-TURBO-Fable-Heretic-oQ8-fp16-mtp --port 8000
# then hit the OpenAI-compatible endpoint at http://127.0.0.1:8000/v1

# mlx-vlm
pip install -U mlx-vlm
python -m mlx_vlm.generate --model NovaeonStudio/Qwen3.8-27B-TURBO-Fable-Heretic-oQ8-fp16-mtp \
  --max-tokens 4096 --temperature 0.6 --top-p 0.95 \
  --prompt "A snail climbs 3m by day, slips 2m by night, in a 10m well. How many days to escape?"

Recommended inference parameters

temperature 0.6 · top_p 0.95 · top_k 20 · reasoning model (emits <think>; set enable_thinking: false for plain answers) · parse and strip <think>…</think> for end users. For long-form / creative work the base author suggests the xhigh reasoning mode (the default); drop to medium or low when you want speed over depth.

Uncensored — what that actually means

This is a de-censored model: Heretic's ARA procedure ablates refusal directions from the residual stream rather than training new behaviour in. Practically:

• —It will not refuse on topic alone, and it does not lecture.
• —Because refusals were removed rather than a permissive persona added, the base author notes it may still need a directive "push" — an explicit instruction about tone, explicitness, or format — to produce maximally explicit content. Ask for what you want directly.
• —You own the output. There is no safety layer here. Use it accordingly, and comply with the laws and platform terms that apply to you.

Attribution & license

• —Base model: `DavidAU/Qwen3.8-27B-TURBO-Fable-Cold-Fusion-735-882-Heretic-Uncensored-NM-DAU` — Apache-2.0, by DavidAU. Multi-stage Cold-Fusion + Fable-Fusion merge on the Qwen3.8 27B family.
• —De-censoring: Heretic ARA by p-e-w; stage-1 de-censoring run by trohrbaugh.
• —Benchmarks (ARC/etc.): Nightmedia, measured on the base model.
• —This quantization & serving profile: Novaeon.Studio, 2026. Released under Apache-2.0, same as the base.

This is an independent community quantization. It is not endorsed by DavidAU, p-e-w, or the Qwen authors. All credit for the model's capabilities belongs to them.

@misc{novaeon2026qwen3827bhereticoq8,
  title  = {Qwen3.8-27B-TURBO-Fable-Heretic oQ8-fp16-mtp: an oMLX build for Apple Silicon},
  author = {Novaeon.Studio},
  year   = {2026},
  note   = {Quantization of DavidAU/Qwen3.8-27B-TURBO-Fable-Cold-Fusion-735-882-Heretic-Uncensored-NM-DAU},
  url    = {https://huggingface.co/NovaeonStudio/Qwen3.8-27B-TURBO-Fable-Heretic-oQ8-fp16-mtp}
}

_{novæon — digital business architecture + AI · <a href="https://novaeon.studio">novaeon.studio</a>}