Qwen3.5-35B-A3B Uncensored (Aggressive) — GGUF Release
A developer's uncensored 35B MoE model achieves 0/465 refusals with full multimodal capabilities intact.
Independent developer HauhauCS has released Qwen3.5-35B-A3B Aggressive, an uncensored version of Alibaba's Qwen3.5 model that completely removes refusal mechanisms while preserving all original capabilities. The model achieved 0 refusals out of 465 test cases with no degradation in performance, maintaining its full multimodal functionality including text, image, and video processing. Available in multiple quantization formats (Q8_0, Q4_K_M, IQ4_XS, etc.) on Hugging Face, the model uses a Mixture-of-Experts architecture with 256 experts and 8+1 active per token, making it efficient for consumer hardware despite its 35B parameter size.
The release represents a significant milestone in the open-source AI community, providing researchers and developers with a powerful, unrestricted tool for specialized applications. HauhauCS worked 12-16 hours daily to ensure high-quality quantization and compatibility with popular inference tools like llama.cpp and LM Studio. The model supports 262K context length and uses hybrid attention with Gated DeltaNet + softmax (3:1 ratio), offering advanced capabilities without the safety filters that typically restrict commercial models. This follows previous aggressive releases in the Qwen3.5 series (4B, 9B, 27B) and continues the trend of community-driven model modifications.
- Achieves 0/465 refusals while maintaining full multimodal (text+image+video) capabilities
- 35B parameter Mixture-of-Experts model with 256 experts, 8+1 active per token for efficiency
- Available in 8 quantization formats on Hugging Face with full vision support via mmproj
Why It Matters
Provides researchers with a fully capable, unrestricted AI model for specialized applications and ethical AI research.