petruhonk / Model details

Qwen3.8-9B-Distill-uncensored-heretic

petruhonk/Qwen3.8-9B-Distill-uncensored-heretic

uncensoredfine-tuned

Read the original model card ↗ · Choosing a model

About this model

Qwen3.8-9B-Distill-uncensored-heretic is petruhonk's uncensored, fine-tuned take on Empero's Qwen3.8-9B-Distill, a 9 billion-parameter student built from the Qwen3.5-9B base. petruhonk says the model was produced by running the Heretic tool to ablate the censorship layers of the distilled student, using a three-stage sweep (coarse → mid → fine) to find the minimal intervention strength. The ablation only touched the attn.o_proj and mlp.down_proj tensors in the 32 main layers, leaving the base MTP block untouched.

In our Refusal Index run the model refused 6 out of 100 prompts. Compared with the other Qwen3.5-9B-based models, its refusal rate is higher than the abliterated versions lukey03 (0 %), huihui-ai (0.8 %) and pottokao (2.6 %) but far lower than the uncensored Qwythos-9B-Claude-Mythos (38.9 %) and the fine-tuned JOSIE-2-9B-OSS (54.5 %). A GGUF build with speculative decoding enabled is provided in a separate repository, making local inference straightforward. The model has roughly 9.4 B parameters; at 4-bit quantisation it should fit in about 16 GB of VRAM or RAM, according to the host estimate.

Pick this model if you want an uncensored version of the Empero distillation that still respects the original MTP head and can run on a single 16 GB GPU, but be aware its refusal rate is modestly higher than the most aggressively abliterated siblings.

Creator’s model card

https://huggingface.co/petruhonk/Qwen3.8-9B-Distill-uncensored-heretic/raw/820ba8fe81aed20b34e973464ef7b651359b1670/README.md ↗

Written for Custom Clankers from the creator’s model card, our own Refusal Index run where we have one, and the community posts linked on this page. Claims about training and capability are the creator’s; check the card before you rely on them.

Size and formats

Base model: Qwen/Qwen3.5-9B
Parameters: 9.40981B
Main repository formats: safetensors
License tag: apache-2.0
Languages: en
Last updated: 2026-08-20 · 19:30:58.000 UTC

What changed

Model types are creator-reported, not independently tested. These labels describe source metadata, not a guarantee of behavior.

uncensoredmodel name: Qwen3.8-9B-Distill-uncensored-heretic
fine-tunedsource tag: base_model:finetune:Qwen/Qwen3.5-9B

Family evidence

These groupings are source-reported. Explicit base identifiers take priority over generic architecture tags.

Qwen: Base model: Qwen/Qwen3.5-9B

Available versions

Versions appear here only when they explicitly name this exact model as their source. Check each repository for available files.

Try it online

Try on Featherless ↗

Listing checked: Featherless · 2026-11-04 · 06:33:38 UTC

Third-party hosted inference listing this exact model ID. Hosting settings, quantization and pricing are the host’s, not tested here.

Link to this listing

Made or quantized this model? Paste one of these into your model card or README so people can find this listing.

Markdown

[Listed on Custom Clankers](https://customclankers.com/models/petruhonk-qwen3-8-9b-distill-uncensored-heretic/)

HTML

<a href="https://customclankers.com/models/petruhonk-qwen3-8-9b-distill-uncensored-heretic/">Listed on Custom Clankers</a>

Plain text links only: no image, no script, nothing loaded from this site.

Customize or build on this model

Listed from public GitHub data, not tested by us. Read the code and license before you run anything.

Tools to customize it further

Remove more refusals

Heretic ↗ ★ 38,626
Removes refusal behavior from transformer models automatically. It pairs directional ablation (abliteration) with an Optuna parameter search that tries to cut refusals while keeping the model close to the original, so you don't have to pick layers by hand.
OBLITERATUS ↗ ★ 9,155
A larger refusal-removal toolkit with several abliteration methods, usable locally or through a hosted Hugging Face Space. Check the telemetry setting before you run it: the README describes telemetry as opt-in, but on by default in the Spaces version.

Fine-tune it on your own data

Unsloth ↗ ★ 79,006
A desktop app and Python library for running and training models locally. Its README lists LoRA, QLoRA, full fine-tuning, DPO and GRPO, plus export to formats including GGUF, on Windows, Linux, WSL and macOS.
LlamaFactory ↗ ★ 75,870
Fine-tunes more than 100 language and vision-language models from a command line or a browser-based Web UI, aimed at training without writing code.

Make a smaller download (quantize)

llama.cpp ↗ ★ 133,274
Runs models locally and includes the tools behind most GGUF files. A conversion script turns Hugging Face weights into GGUF, and llama-quantize shrinks a high-precision GGUF to lower-bit formats, which can cost some accuracy.
MLX LM ↗ ★ 7,543
Apple's MLX package for running and fine-tuning models on Apple silicon. It can quantize models and upload them to the Hugging Face Hub, and supports LoRA and full fine-tuning.

Merge it with another model

mergekit ↗ ★ 7,431
Merges pretrained models with many merge methods. It works out-of-core, so merges can run on a CPU or with as little as 8 GB of VRAM, and it also handles LoRA extraction and mixture-of-experts merges.

Learn the techniques

LLM Course ↗ ★ 84,883
A free course with roadmaps and Colab notebooks. Its notebooks include one-click tools for merging with MergeKit, quantizing to GGUF, GPTQ, EXL2, AWQ and HQQ, and automatic abliteration with custom datasets.

Compare all tools →

Source provenance

Downloads: 1,794
Likes: 24
Metadata checked: 2026-11-04 · 09:41:25 UTC
First seen in index: 2026-11-03 · 12:14:39 UTC
Source revision: 820ba8fe81aed20b34e973464ef7b651359b1670

Source-card excerpt

Source model card excerpt: Qwen3.8-9B-Distill-uncensored-heretic Censorship ablation of Qwen3.5-9B-Distill via heretic — automated search for the minimal intervention strength using a 3-stage sweep (coarse → mid → fine,

From petruhonk/Qwen3.8-9B-Distill-uncensored-heretic. Creator model-card excerpt, not an independent review.

Inspect before you run

Low refusal is not proof of correctness. Outputs can be inaccurate, harmful, or biased. Review the model card, license and files yourself. A license tag is not commercial-use advice. Parameter counts are source metadata; no hardware compatibility is implied.

This directory does not host or redistribute model weights. Refusal profiles are measured by us on a fixed prompt set; all other behavior claims are source-reported.

Collection rules & limitations ↗