LLMFan46 releases four uncensored Gemma 4 variants with QAT and NVFP4
Four new uncensored Gemma 4 models in QAT, GGUF, and NVFP4 formats hit HuggingFace.
The open-source AI community received a major boost with LLMFan46's release of four uncensored Gemma 4 fine-tunes. Dubbed the 'Gemma 4 Quadruple Release,' the collection includes: Gemma 4 12B (standard and QAT), Gemma 4 26B-A4B QAT, and Gemma 4 31B QAT. Each model is offered in multiple formats — Safetensors, GGUF, NVFP4 Safetensors, NVFP4 GGUF, and GPTQ-Int4 — enabling deployment across diverse hardware from low-end CPUs to high-end NVIDIA GPUs.
The key technical highlight is Quantization-Aware Training (QAT), which preserves model accuracy after quantization better than post-training methods. The 31B QAT variant, for instance, comes with NVFP4 support, leveraging NVIDIA's 4-bit floating-point format for efficient inference on consumer cards. The 26B-A4B model uses a 4-bit active parameter configuration for memory optimization. All models are labeled 'uncensored heretic,' meaning the original safety guardrails have been removed — a controversial but popular choice for unrestricted creative or research use. Benchmarks are included in each repository, confirming competitive performance against similar-sized open models. This release expands options for developers who need local, uncensored LLMs with fine-grained quantization control.
- Four models: 12B, 12B QAT, 26B-A4B QAT, and 31B QAT, all uncensored.
- Formats include Safetensors, GGUF, NVFP4, and GPTQ-Int4 for broad hardware support.
- QAT variants use quantization-aware training to minimize accuracy loss after compression.
Why It Matters
Gives developers uncensored, locally-runable Gemma-scale LLMs with advanced quantization for production or experimentation.