akpsahan/Gemma4-26B-A4B-QAT-Uncensored-HauhauCS-Balanced-MTP

Gemma4-26B-A4B-QAT-Uncensored-HauhauCS-Balanced-MTP

[Join the Discord](https://discord.gg/SZ5vacTXYf) for updates, roadmaps, projects, or just to chat.

Gemma4-26B-A4B (QAT) uncensored by HauhauCS. 0/465 Refusals*

About

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.

Balanced

The Balanced variant (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. An Aggressive variant, for cases where Balanced still deflects too much, after current testing is not required.

~35% faster with MTP

Ships with an MTP (multi-token-prediction) draft head for speculative decoding — roughly 35% 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 Gemma4-26B-A4B-QAT-Uncensored-HauhauCS-Balanced-Q4_K_M.gguf \
  -md mtp-gemma-4-26B-A4B-it.gguf --spec-type draft-mtp \
  -ngl 99 -fa on

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

Downloads

Why only Q4_K_M? Gemma 4 is quantization-aware-trained for ~4-bit, so Q4KM is the sweet spot — higher-precision quants add size with no real quality gain. Carefully quantized for best quality at 4-bit.

Vision

Load the mmproj alongside the model for image input:

llama-server -m Gemma4-26B-A4B-QAT-Uncensored-HauhauCS-Balanced-Q4_K_M.gguf \
  --mmproj mmproj-Gemma4-26B-A4B-QAT-Uncensored-HauhauCS-Balanced-BF16.gguf -ngl 99 -fa on

Recommended 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.

Specs

• —26B-A4B MoE (128 experts, 8 active per token) · 256K (262144) context
• —Vision (image input) via mmproj
• —Based on Gemma 4 26B-A4B by Google DeepMind

Compatibility

• —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's tensor-split mode — use a single GPU (layer-split or priority order) for this model.

Acknowledgements

• —Google DeepMind — Gemma 4.
• —The included mtp-gemma-4-26B-A4B-it.gguf speculative draft head comes from Unsloth'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 Discord](https://discord.gg/SZ5vacTXYf) and flag it so I can work on it in a future revision.