Quantization	File	Size
Q8_0	Qwen3-VL-8B-Instruct-heretic-Q8_0.gguf	8.11 GB
Q6_K	Qwen3-VL-8B-Instruct-heretic-Q6_K.gguf	6.26 GB
Q4_K_M	Qwen3-VL-8B-Instruct-heretic-Q4_K_M.gguf	4.68 GB

Usage with llama.cpp (Recommended)

Note: Ollama (as of v0.16.x) has a known bug that crashes when loading Qwen3-VL models. Use llama.cpp directly for vision features.

Vision models require a separate multimodal projector (mmproj) file. Download the official mmproj from Qwen/Qwen3-VL-8B-Instruct-GGUF:

bash

# Download mmproj
huggingface-cli download Qwen/Qwen3-VL-8B-Instruct-GGUF mmproj-Qwen3VL-8B-Instruct-F16.gguf

# Run with llama-server (OpenAI-compatible API)
llama-server \
  -m Qwen3-VL-8B-Instruct-heretic-Q8_0.gguf \
  --mmproj mmproj-Qwen3VL-8B-Instruct-F16.gguf \
  -ngl 999

# Or use the CLI directly
llama-mtmd-cli \
  -m Qwen3-VL-8B-Instruct-heretic-Q8_0.gguf \
  --mmproj mmproj-Qwen3VL-8B-Instruct-F16.gguf \
  --image photo.jpg \
  -p "Describe this image." \
  -ngl 999

Usage with Ollama (Text Only)

Ollama can load this model for text-only chat, but vision/image features will crash due to the bug linked above.

bash

ollama run hf.co/ThalisAI/Qwen3-VL-8B-Instruct-heretic:Q8_0
ollama run hf.co/ThalisAI/Qwen3-VL-8B-Instruct-heretic:Q6_K
ollama run hf.co/ThalisAI/Qwen3-VL-8B-Instruct-heretic:Q4_K_M

bf16 Weights

The full bf16 abliterated weights are available in the bf16/ subdirectory of this repository.

Usage with Transformers

The bf16 weights in the bf16/ subdirectory can be loaded directly with Transformers:

python

from transformers import AutoModelForImageTextToText, AutoTokenizer

model_id = "ThalisAI/Qwen3-VL-8B-Instruct-heretic"
tokenizer = AutoTokenizer.from_pretrained(model_id, subfolder="bf16")
model = AutoModelForImageTextToText.from_pretrained(
    model_id, subfolder="bf16", torch_dtype="auto", device_map="auto"
)

messages = [{"role": "user", "content": "Hello!"}]
text = tokenizer.apply_chat_template(messages, add_generation_prompt=True, tokenize=False)
inputs = tokenizer(text, return_tensors="pt").to(model.device)
outputs = model.generate(**inputs, max_new_tokens=512)
print(tokenizer.decode(outputs[0][inputs.input_ids.shape[1]:], skip_special_tokens=True))

About

This model was processed by the Apostate automated abliteration pipeline:

The source model was loaded in bf16
Heretic's optimization-based abliteration was applied to remove refusal behavior
The merged model was converted to GGUF format using llama.cpp
Multiple quantization levels were generated

The abliteration process uses directional ablation to remove the model's refusal directions while minimizing KL divergence from the original model's behavior on harmless prompts.