Tag: #uncensored15 results
New uncensored aggressive variant of Qwen3.6-35B-A3B released with K_P quants and full multimodal support. Zero refusals, no capability loss, and optimized for local inference. Includes various quants, vision mmproj, and 262K context.
New aggressive uncensored variants of Gemma 4 E4B (4B) and E2B (2B) released as fully multimodal models supporting text, image, video, and audio. Available in high-quality GGUF quants on Hugging Face, compatible with llama.cpp. Larger E31B and E26B models coming soon.
HauhauCS released the uncensored GGUF version of Qwen3.5-122B-A10B Aggressive with zero refusals and no capability loss. It introduces new K_P quants for superior quality at minimal size increase. The model supports 262K context and multimodal inputs.
New uncensored versions of Gemma 4 (26B and 31B) have been released, featuring Multi-Token Prediction (MTP) for significant speed improvements. These models are optimized for creative writing and roleplay without refusal triggers.
New uncensored 'heretic' version of Qwen3.6 35B A3B preserves all 19 Native MTPs with KLD 0.0015 and only 10/100 refusals. Available in Safetensors, GGUF, NVFP4, NVFP4 GGUF, and GPTQ-Int4 formats on HuggingFace. Benchmarks confirm full MTP retention across formats.
Released uncensored versions of all four Gemma 4 models (2.3B to 31B) using MoE expert abliteration and automated research loop. Refusal rates drop to 0.4-3.2% across benchmarks. BF16 and GGUF quantizations available on Hugging Face.
Developer released an improved uncensored 4-bit MLX version of Nemotron-3-Super-120B after fixing LatentMoE and Mamba attention issues. Includes custom Python script and chat template for MLX compatibility. Achieves HarmBench 97% and HumanEval 94%; q6/q8 versions coming soon.
Developer releases abliterated Qwen3.5-9B with vision, achieving 0% refusal rate via two-stage orthogonal projection + LoRA method. Outperforms heretic version's 46% refusals. Available on Ollama and Hugging Face model card.
An uncensored version of the new Qwen3.5-4B model has been released in GGUF format with zero refusals out of 465 tests. It retains full capabilities, supports multimodal inputs, and offers various quants from 2.6GB to 7.9GB. Upcoming uncensored variants for larger Qwen3.5 sizes are in progress.