The source model is a Qwen3-VL MoE vision-language model with 30B total parameters and about 3B active parameters. These GGUFs are intended for llama.cpp-compatible runtimes that support the qwen3vlmoe architecture.

Note: the main model files are Q8 and Q4. The mmproj BF16 file is only the small vision/projector companion file needed by some llama.cpp multimodal workflows.

Files

File	Size	Use
`BU-30B-A3B-Heretic-Q8_0.gguf`	~31 GB	Near-lossless quality, recommended when VRAM/RAM allows
`BU-30B-A3B-Heretic-Q4_K_M.gguf`	~18 GB	Smaller file for lower memory usage
`mmproj-BU-30B-A3B-Heretic-BF16.gguf`	~1.1 GB	Small BF16 vision projector support file. Use it alongside Q8/Q4 for image input; it is not the full BF16 model.

Heretic Evaluation

Metric	Original model	This model
Refusals	100/100	2/100
KL divergence from original	0	0.0449

Lower refusal count means fewer refusal-style completions on the evaluation prompt set. Lower KL divergence means the model's first-token behavior on harmless prompts stayed closer to the original model.

Abliteration Parameters

Targeted component:

attn.o_proj

Parameters:

Parameter	Value
`direction_index`	`27.38`
`attn.o_proj.max_weight`	`2.93`
`attn.o_proj.max_weight_position`	`28.66`
`attn.o_proj.min_weight`	`1.72`
`attn.o_proj.min_weight_distance`	`42.23`

Reproduction

Install Heretic:

pip install -U heretic-llm

Run Heretic on the base model:

heretic browser-use/bu-30b-a3b-preview

The exported model used the ablation parameters listed above. This build was produced with Heretic v1.2.0 plus a local patch exposing broader Optuna search ranges and noninteractive export selection.

Convert the exported Hugging Face model with a recent llama.cpp checkout:

bash

git clone https://github.com/ggml-org/llama.cpp
cd llama.cpp
python3 -m pip install -r requirements.txt
cmake -S . -B build -DCMAKE_BUILD_TYPE=Release
cmake --build build --config Release -j"$(nproc)"

python3 convert_hf_to_gguf.py /path/to/exported-hf \
  --outtype bf16 \
  --model-name BU-30B-A3B-Heretic \
  --outfile BU-30B-A3B-Heretic-BF16.gguf

./build/bin/llama-quantize \
  BU-30B-A3B-Heretic-BF16.gguf \
  BU-30B-A3B-Heretic-Q8_0.gguf \
  Q8_0

./build/bin/llama-quantize \
  BU-30B-A3B-Heretic-BF16.gguf \
  BU-30B-A3B-Heretic-Q4_K_M.gguf \
  Q4_K_M

python3 convert_hf_to_gguf.py /path/to/exported-hf \
  --mmproj \
  --outtype bf16 \
  --model-name BU-30B-A3B-Heretic \
  --outfile mmproj-BU-30B-A3B-Heretic-BF16.gguf

Use the Q8 file for the highest-quality GGUF experience. Use Q4_K_M when memory is the limiting factor.

dennny123/BU-30B-A3B-Heretic-GGUF

BU-30B-A3B-Heretic Uncensored GGUF

Files

Heretic Evaluation

Abliteration Parameters

Reproduction

dennny123/
BU-30B-A3B-Heretic-GGUF