HauhauCS releases uncensored Gemma4 QAT models with MTP: up to 53% speed boost
New Gemma4 builds offer 35-53% faster inference via multi-token prediction and zero refusals on edge cases.
HauhauCS, a prolific developer on Hugging Face nearing 20M downloads, has released two new uncensored Gemma4 QAT models: Gemma4-26B-A4B-QAT-Uncensored-Balanced-MTP and Gemma4-31B-QAT-Uncensored-Balanced-MTP. Both are based on Google's Gemma-4 and include a multi-token-prediction (MTP) draft head for speculative decoding, which yields 35% faster inference on the 26B-A4B (a Mixture-of-Experts model with 128 experts, 8 active per token, ~4B active parameters) and 53% faster on the dense 31B model. The speedup comes without quality loss because the model verifies each drafted token. The 'Balanced' variant adds a light reasoning preamble for the most extreme prompts before delivering a full answer, achieving 0 out of 465 refusals in GenRM tests. No aggressive variant is needed. The models are optimized for creative writing, role-play, and emotional intelligence—though for agentic coding/tool use, HauhauCS recommends Qwen3.6 instead.
Key technical details: both models support 262K context and vision (via mmproj). The 26B-A4B Q4_K_M is 16.8 GB, the 31B is 18.7 GB. Recommended sampling parameters are temp=0.6, top_k=64, top_p=0.9, min_p=0.05, repeat_penalty=1.1. In llama.cpp, enable MTP via flags like `-md mtp-gemma-4-26B-A4B-it.gguf --spec-type draft-mtp`. The developer notes stable sampling, no looping, and long-context coherence. Users should be aware that Gemma 4 can crash under LM Studio's tensor-split mode on multi-GPU setups. These models are recommended for users who want unfiltered, high-performance local AI generation.
- MTP speculative decoding provides 35% speedup on 26B-A4B (MoE, 4B active params) and 53% on dense 31B, with zero quality loss.
- Balanced uncensored variant passes 0/465 refusals in GenRM tests while keeping original model personality and coherence.
- Models support 262K context, vision, and are optimized for Q4_K_M quantization (16.8-18.7 GB) – the quality sweet spot due to QAT training.
Why It Matters
Professionals get fast, uncensored local inference for creative tasks, with MTP setting a new standard for efficient LLM deployment.