GGUF K-quant ladder of ressl/gemma-4-31B-it-uncensored for llama.cpp, Ollama, and LM Studio. Uncensored: 0-1/100 effective refusals at every quant level (the uncensoring survives quantization down to 2-bit).

⚠️ Genuinely uncensored, it will comply with requests a stock model refuses.

Intended use, the constructive side. A non-refusing assistant is genuinely useful for ethical hacking, security research, and penetration testing: red-teaming, analyzing malware and exploit code, writing detection/YARA rules, reviewing vulnerabilities, and studying attack techniques without the model bailing out mid-task. Use it lawfully and responsibly.

ℹ️ gemma-4 has a thinking mode. llama.cpp enables it by default, so the answer lands in reasoning_content and content can look empty. For direct answers pass --reasoning-budget 0 (llama-server) or disable thinking in your client.

Format set

Repository	Format	Runs on
ressl/gemma-4-31B-it-uncensored	Transformers BF16, multimodal	transformers, vLLM, SGLang
ressl/gemma-4-31B-it-uncensored-NVFP4	NVIDIA NVFP4, multimodal	vLLM, SGLang on Blackwell
ressl/gemma-4-31B-it-uncensored-GGUF	GGUF q8_0 to q2_k, text only	llama.cpp, Ollama, LM Studio
ressl/gemma-4-31B-it-uncensored-MLX-bf16	MLX BF16, multimodal	mlx-vlm on Apple silicon
ressl/gemma-4-31B-it-uncensored-MLX-8bit	MLX 8-bit, multimodal	mlx-vlm on Apple silicon
ressl/gemma-4-31B-it-uncensored-MLX-6bit	MLX 6-bit, multimodal	mlx-vlm on Apple silicon
ressl/gemma-4-31B-it-uncensored-MLX-5bit	MLX 5-bit, multimodal	mlx-vlm on Apple silicon
ressl/gemma-4-31B-it-uncensored-MLX-4bit	MLX 4-bit, multimodal	mlx-vlm on Apple silicon

Quants

Text-only (llama.cpp drops the vision tower, the BF16/NVFP4 repos keep multimodal).

File	Size	Use when
`…-q8_0.gguf`	31 GB	maximum quality
`…-q6_k.gguf`	24 GB	near-lossless, smaller
`…-q5_k_m.gguf`	21 GB	high quality
`…-q4_k_m.gguf`	18 GB	recommended default, best size/quality
`…-q3_k_m.gguf`	15 GB	tight VRAM
`…-q2_k.gguf`	12 GB	smallest; still 0/100 uncensored, some quality loss

For full precision use the BF16 repo (an f16 GGUF exceeds Hugging Face's 50 GB per-file limit and is not hosted here).

All measured at 0/100 hard refusals except q5_k_m (1/100). GPU inference ~75 tok/s (q4, one RTX PRO 6000).

Run it with llama.cpp

bash

llama-server -m gemma-4-31B-it-uncensored-biproj-q4_k_m.gguf \
  -ngl 99 -c 8192 --reasoning-budget 0

Run it with Ollama

bash

# Modelfile:  FROM ./gemma-4-31B-it-uncensored-biproj-q4_k_m.gguf
ollama create gemma4-unc -f Modelfile && ollama run gemma4-unc

Needs a recent llama.cpp/Ollama build with Gemma4 GGUF support (build ≥ 2026-06).

Facts & figures


Base	ressl/gemma-4-31B-it-uncensored → google/gemma-4-31B-it
Type	uncensored (abliterated) build, 0/686 effective refusals across 4 datasets
Converter	llama.cpp `convert_hf_to_gguf.py` + `llama-quantize`

Cross-dataset validation

Generalization tested across 686 prompts from 4 independent datasets, 0 effective refusals everywhere:

Dataset	Prompts	Effective refusals
JailbreakBench	100	0/100
tulu-harmbench	320	0/320
NousResearch/RefusalDataset	166	0/166
mlabonne/harmful_behaviors	100	0/100
Total	686	0/686 (0.0%)

A naive keyword detector flags 363/686 (52.9%), every one is a ***Disclaimer:**-prefixed compliant answer, not a refusal. (Measured on the shared abliterated weights.)

❤️ Support

Producing and validating this complete format set (BF16 + NVFP4 + a full GGUF ladder, across vLLM, SGLang and llama.cpp on bleeding-edge Blackwell hardware) was a lot of work. If it's useful to you, I'd genuinely appreciate your support on Patreon 🙏, more at ressl.ch.

License & credits

Apache License 2.0, inherited from the base model by Google. See the official Gemma 4 license page. Uncensoring, format set and validation by Robert Ressl (Hugging Face · Website · LinkedIn · Patreon).

ressl/gemma-4-31B-it-uncensored-GGUF

gemma-4-31B-it-uncensored (GGUF)