HauhauCS/Gemma4-12B-QAT-Uncensored-HauhauCS-Balanced · Hugging Face

Source: https://huggingface.co/HauhauCS/Gemma4-12B-QAT-Uncensored-HauhauCS-Balanced

https://huggingface.co/HauhauCS/Gemma4-12B-QAT-Uncensored-HauhauCS-Balanced#gemma4-12b-qat-uncensored-hauhaucs-balancedGemma4-12B-QAT-Uncensored-HauhauCS-Balanced

**Join the Discord**for updates, roadmaps, projects, or just to chat.

Gemma4-12B (QAT) uncensored by HauhauCS.0/465 Refusals*

https://huggingface.co/HauhauCS/Gemma4-12B-QAT-Uncensored-HauhauCS-Balanced#aboutAbout

No changes to datasets or capabilities — fully functional, 100% of what the original authors intended, just without the refusals. Built from the official QAT weights, so the 4-bit quant stays close to full-precision quality.

https://huggingface.co/HauhauCS/Gemma4-12B-QAT-Uncensored-HauhauCS-Balanced#balancedBalanced

TheBalancedvariant (recommended — 99%+ of users will be happy here) uses optimized full uncensoring tuned especially for agentic coding, reasoning, creative writing and reliability-critical tasks. It reasons before answering and stays dependable and on-instruction. AnAggressivevariant, for cases where Balanced still deflects too much, after current testing is not required.

https://huggingface.co/HauhauCS/Gemma4-12B-QAT-Uncensored-HauhauCS-Balanced#60-faster-with-mtp~60% faster with MTP

Ships with an MTP (multi-token-prediction) draft head forspeculative decoding— roughly60% faster generation with identical output(the model verifies every drafted token, so quality is unchanged — pure speed). This release is tuned to pair well with the included MTP head.

llama.cpp:

llama-server \
  -m Gemma4-12B-QAT-Uncensored-HauhauCS-Balanced-Q4_K_M.gguf \
  -md mtp-gemma-4-12B-it.gguf --spec-type draft-mtp \
  -ngl 99 -fa on

Note:the MTP speedup was currently tested by me throughllama.cpp(llama\-server/llama\-cli).

https://huggingface.co/HauhauCS/Gemma4-12B-QAT-Uncensored-HauhauCS-Balanced#downloadsDownloads

FileTypeSizeGemma4\-12B\-QAT\-Uncensored\-HauhauCS\-Balanced\-Q4\_K\_M\.ggufQ4_K_M (text)6.9 GBmmproj\-Gemma4\-12B\-QAT\-Uncensored\-HauhauCS\-Balanced\-BF16\.ggufmmproj (vision)168 MBmtp\-gemma\-4\-12B\-it\.ggufMTP speculative drafter242 MB

**Why only Q4_K_M?**Gemma 4 is quantization-aware-trained for ~4-bit, so Q4_K_M is the sweet spot — higher-precision quants add size with no real quality gain. Carefully quantized for best quality at 4-bit.

https://huggingface.co/HauhauCS/Gemma4-12B-QAT-Uncensored-HauhauCS-Balanced#visionVision

Load the mmproj alongside the model for image input:

llama-server -m Gemma4-12B-QAT-Uncensored-HauhauCS-Balanced-Q4_K_M.gguf \
  --mmproj mmproj-Gemma4-12B-QAT-Uncensored-HauhauCS-Balanced-BF16.gguf -ngl 99 -fa on

https://huggingface.co/HauhauCS/Gemma4-12B-QAT-Uncensored-HauhauCS-Balanced#recommended-samplingRecommended sampling

These are dialed in specifically for this HauhauCS build — use them for the intended behaviour and quality:

• temperature 0\.6
• top\_k 64
• top\_p 0\.9
• min\_p 0\.05
• repeat\_penalty 1\.1

This release is tuned end-to-end as its own thing; the settings above are part of that and aren’t the stock Gemma defaults.

https://huggingface.co/HauhauCS/Gemma4-12B-QAT-Uncensored-HauhauCS-Balanced#specsSpecs

• 12B dense · 256K (262144) context
• Vision (image input) via mmproj
• Based onGemma 4 12Bby Google DeepMind

https://huggingface.co/HauhauCS/Gemma4-12B-QAT-Uncensored-HauhauCS-Balanced#compatibilityCompatibility

• Works with llama.cpp, LM Studio, Jan, koboldcpp, and other GGUF runtimes.
• **Multi-GPU + LM Studio:**I’ve personally noticed Gemma 4 can crash under LM Studio’stensor-splitmode — use a single GPU (layer-split or priority order) for this model.

https://huggingface.co/HauhauCS/Gemma4-12B-QAT-Uncensored-HauhauCS-Balanced#acknowledgementsAcknowledgements

• Google DeepMind— Gemma 4.
• The includedmtp\-gemma\-4\-12B\-it\.ggufspeculative draft head comes fromUnsloth’s Gemma 4 release — many thanks to the Unsloth team for it.

*Tested with both automated and manual refusal benchmarks — none have been found in standard use. A small number of edge-case prompts deflect on the first ask but comply on a re-ask or strategic framing. If you hit one that’s actually obstructive to your use case,join the Discordand flag it so I can work on it in a future revision.