Skip to content

Search wahloon.net

Search models, news, guides and FAQs.

All news
News#qwen#27b#abliterated#heretic#roundup

Qwen3.8-27B abliterated builds: a practical roundup

Five popular Qwen3.8-27B abliterated and Heretic variants, what separates them, and which one to try first on a 24 GB card.

3 min readMarkdown
On this page

Qwen3.8-27B has become the default mid-size base for abliterated builds. It fits on a 24 GB GPU at Q4, has a permissive license, and the community has produced several distinct variants. The differences are mostly method, quantization coverage, and how aggressively refusals were removed.

Here is a practical comparison of the builds currently in the catalog. Download counts are approximate 30-day figures from the data file and change constantly.

The builds at a glance

BuildMethodApprox. VRAMFormatsNotes
Huihui-Qwen3.8-27B-abliteratedAbliteration~16–18 GBGGUF, Transformers, NVFP4Most downloaded; broad format coverage
Qwen3.8-27B-Uncensored (JonathanColetti)Abliteration~16.5 GBGGUF, AWQ, FP8, TransformersMultiple serving formats
HauhauCS Aggressive MTPFine-tune~16.5 GBGGUFAggressive low-refusal tuning
0bserverx Heretic-ARAHeretic-ARA~16–17 GBGGUFRefined, low KL divergence claims
DavidAU TURBO FableHeretic~16.5 GBGGUF, FP8, TransformersInstruction-following focus

Huihui: the safe default

The Huihui build is the one most people should start with. It is widely mirrored, available in GGUF and Transformers formats, and has an Ollama tag:

ollama run huihui_ai/Qwen3.8-abliterated:27b

It is plain abliteration rather than a multi-stage Heretic pipeline, so behavior stays close to the base model. If you want the smallest number of surprises, this is the least exotic option.

JonathanColetti: format flexibility

This variant adds AWQ and FP8 alongside GGUF and Transformers. That matters if you serve with vLLM or another engine that prefers AWQ/FP8 over GGUF. For a single local chat, there is no strong reason to pick it over Huihui; for a serving stack, the extra formats can simplify deployment.

HauhauCS: aggressive fine-tune

The HauhauCS build is a fine-tune rather than a pure weight edit. It is tuned for very low refusal rates, and that shows in how it handles borderline prompts. The trade-off is the usual one: aggressive tuning can drift further from the base model's style and instruction-following behavior.

Try it when the standard abliterated build still hedges too much for your use case. Test it on your actual tasks before standardizing on it.

0bserverx: Heretic-ARA refinement

This is the most “processed” option in the group. Heretic automates the ablation search, and ARA adds a refinement pass intended to reduce residual refusals while limiting capability drift. The model card emphasizes low refusal scores and low KL divergence.

That combination is appealing if you want low refusals without the rough edges of an aggressive fine-tune. As always, the claims are worth verifying on your own prompts.

DavidAU: TURBO Fable

DavidAU's TURBO Fable line targets strong instruction following with reduced refusals. It is available in GGUF, FP8, and Transformers formats, and is a reasonable choice if you want a Heretic-style build with more serving options.

Which should you pick?

  • First local 27B: Huihui. Broad support, predictable behavior.
  • Serving with vLLM: JonathanColetti for AWQ/FP8, or DavidAU for FP8.
  • Lowest refusals: HauhauCS, with the caveat that it is a fine-tune.
  • Balanced low-refusal edit: 0bserverx Heretic-ARA.

All five are 27B dense models. At Q4 they need roughly 16–18 GB of VRAM, and more if you raise context. On a 16 GB card, use a Q3/IQ3 quant or a smaller base model.

A note on numbers

Download counts and likes are noisy signals. A high count can mean quality, good discoverability, or simply that a build was first. Check the model card, the quantization list, and the last-updated date before committing.

The catalog stores these figures in data/models.ts so they can be refreshed without touching the UI. If a build is missing or a number is stale, use the contribute page to suggest an update.

Next steps

Model behavior changes; licenses and law do not. Verify the current model card before downloading, and note that you are responsible for how you use a model and for the content you generate with it.

3 minNews

MoE abliteration arrives for Qwen3.6-35B-A3B

The 35B-A3B MoE base is now the most interesting size/quality trade-off for local inference. Here are the abliterated and uncensored builds worth testing.

#qwen#moe#35b-a3b
Read