ethanfel / Model details

Qwen3-VL-32B-Ultra-Heretic-H3-ComfyUI-INT8-ConvRot

ethanfel/Qwen3-VL-32B-Ultra-Heretic-H3-ComfyUI-INT8-ConvRot

abliterateduncensoredfine-tuned

Read the original model card ↗ · Choosing a model

About this model

ethanfel's ComfyUI-oriented repackaging of llmfan46/Qwen3-VL-32B-Instruct-ultra-uncensored-heretic, provided as safetensors weights under an Apache-2.0 license. The repository contains H3 conditioning encoders with Qwen3-VL language layers 0–49 and the full vision tower in BF16 and INT8 ConvRot formats, plus optional generation-only tails with layers 50–63, the final norm, and LM head in BF16, INT8 ConvRot, and NVFP4/AWQ. The model card says the encoder intentionally omits layers 50–63, the final language norm, and the LM head, and that these are ComfyUI checkpoints rather than a complete Transformers generation repository. According to the card, the upstream source's Heretic v1.2.0 ARA edits target attention output projections in language layers 31–40, all within the retained 0–49 range, so those uncensoring edits are present; the card also notes abliteration reduces refusals but does not guarantee removal of all safety behavior and may affect quality.

Description sources

https://huggingface.co/ethanfel/Qwen3-VL-32B-Ultra-Heretic-H3-ComfyUI-INT8-ConvRot/raw/e8967f6a39ea5b4939a1aff81be3e8706490c0e8/README.md ↗

Source-based description, not an independent evaluation. Check the linked sources for the creator’s claims and limitations.

Size and formats

Base model: llmfan46/Qwen3-VL-32B-Instruct-ultra-uncensored-heretic
Parameters: Unknown
Main repository formats: safetensors
License tag: apache-2.0
Languages: en
Last updated: 2026-08-05 · 15:28:00.000 UTC

Community reviews & tips

Source-linked paraphrases of sampled community sources, not independently tested by this directory.

Sample window: 2025-10-24 UTC to 2026-10-24 UTC · 3 unique source links. This is a selected sample, not all community discussions.

Small sample: source links are not a count of independent reviewers or evidence of consensus.

Sources checked: 2026-10-24 22:53:16 UTC

Hands-on report
- One user reports successfully running the model on a 5070 Ti and generating videos, with a 15-second 640×832 video taking about 20 minutes.
Tested revision: Unknown · Quantization: Unknown
- Hugging Face discussion · 2026-08-06 UTC ↗
Practical tip
- The repo owner suggests using the CLI to download the model, saying it is much faster; another user still reports very slow download speeds.
Tested revision: Unknown · Quantization: Unknown
- Hugging Face discussion · 2026-08-29 UTC ↗
Practical tip
- One user explains that this model follows the orthodox Minimax prompt structure but, when used as a text encoder, will not reformat prompts; they advise writing prompts by hand or using a separate local multimodal LLM as a prompt engineer, and note debate about whether heretic models may worsen Minimax results.
Tested revision: Unknown · Quantization: Unknown
- Hugging Face discussion · 2026-08-10 UTC ↗

What changed

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

abliteratedsource tag: abliterated
uncensoredsource tag: base_model:finetune:llmfan46/Qwen3-VL-32B-Instruct-ultra-uncensored-heretic
uncensoredsource tag: base_model:llmfan46/Qwen3-VL-32B-Instruct-ultra-uncensored-heretic
uncensoredsource tag: uncensored
fine-tunedsource tag: base_model:finetune:llmfan46/Qwen3-VL-32B-Instruct-ultra-uncensored-heretic

Family evidence

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

Qwen: Base model: llmfan46/Qwen3-VL-32B-Instruct-ultra-uncensored-heretic

Available versions

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

Main repository files ↗safetensors
No other explicitly linked versions recorded.

Refusal profile

Not yet measured: no GGUF version is available for our runner.

How we measure

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/ethanfel-qwen3-vl-32b-ultra-heretic-h3-comfyui-int8-convrot/)

HTML

<a href="https://customclankers.com/models/ethanfel-qwen3-vl-32b-ultra-heretic-h3-comfyui-int8-convrot/">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 ↗ ★ 36,421
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 ↗ ★ 8,946
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 ↗ ★ 78,528
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 ↗ ★ 75,713
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 ↗ ★ 132,260
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 ↗ ★ 7,467
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 ↗ ★ 7,414
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 ↗ ★ 84,667
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: 0
Likes: 731
Metadata checked: 2026-10-24 · 04:11:48 UTC
First seen in index: 2026-10-24 · 04:11:48 UTC
Source revision: e8967f6a39ea5b4939a1aff81be3e8706490c0e8

Source-card excerpt

Source model card excerpt: Qwen3-VL-32B Ultra Uncensored Heretic — H3 ComfyUI encoders + generation tails This repository contains ComfyUI H3 conditioning encoders and optional Qwen3-VL-32B generation tails. The

From ethanfel/Qwen3-VL-32B-Ultra-Heretic-H3-ComfyUI-INT8-ConvRot. 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 ↗