Qwen3.6-35B-A3B Uncensored Aggressive
HauhauCS/Qwen3.6-35B-A3B-Uncensored-HauhauCS-Aggressive
published Apr 2026 · updated Apr 2026
Qwen3.6-35B-A3B Uncensored Aggressive is a vlm model that removes refusals from the original Qwen3.6-35B-A3B while preserving full functionality and multimodal capabilities.
specs
| Task | Multimodal language model (text, image, video) |
| Architecture | Mixture of Experts (MoE) with 256 experts, 8 routed per token; hybrid linear + softmax attention (3:1 ratio) |
| Parameters | 35B total, ~3B active per forward pass |
| License | Unspecified (based on Qwen/Qwen3.6-35B-A3B) |
about this model
Qwen3.6-35B-A3B-Uncensored-HauhauCS-Aggressive is a vision-language model (VLM) that combines the full multimodal and reasoning capabilities of the base Qwen3.6-35B-A3B with lossless uncensoring. Based on Qwen/Qwen3.6-35B-A3B, it uses a Mixture-of-Experts architecture (35B total parameters, ~3B active per forward pass, 256 experts, 8 routed per token) with hybrid linear and full softmax attention (3:1 ratio). It supports text, image, and video input with a native 262K-token context window.
Key Strengths
- Zero refusals: Tested against a set of 465 prompts, the model shows 0 refusals across all content categories.
- Aggressive uncensoring variant: The model is fully unlocked and will not refuse prompts. Short disclaimers baked into the base model training may occasionally appear, but full content is always generated.
- Optimized quantization: Hosted quants are generated with importance-matrix (imatrix) calibration on abliterated weights. Custom K_P (“Perfect”) quantizations preserve quality at 1–2 levels above the base quant with only ~5–15% additional file size, compatible with any GGUF runtime.
Architecture Summary
| Property | Value |
|---|---|
| Total parameters | 35B |
| Active parameters per token | ~3B |
| Number of experts | 256 (8 routed) |
| Attention type | Hybrid (linear + full softmax, 3:1 ratio) |
| Layers | 40 |
| Context length | 262K tokens |
| Modality | Text, image, video |
Recommended inference settings (from the base Qwen authors) are available in the model card for both thinking and non-thinking modes. Vision support requires a multimodal projection file (mmproj) alongside the GGUF, which is included in the hosted serving stack.
best for
- ·Uncensored creative writing and roleplay without content filters
- ·Multimodal reasoning over images and video with no refusal restrictions
FAQ
It applies stronger uncensoring so the model will not refuse any prompt, though it may occasionally append short disclaimers baked into the base training.
It supports 262K native context with a hybrid attention mechanism (3:1 ratio of linear to full softmax) and 256 experts with 8 routed per token.
Use the gigarouter OpenAI-compatible endpoint with your API key, specifying the model name Qwen3.6-35B-A3B-Uncensored-HauhauCS-Aggressive.
It accepts text, image, and video inputs natively; the mmproj file must be loaded alongside the GGUF for vision tasks.
For thinking mode: temperature=1.0, top_p=0.95, top_k=20, presence_penalty=1.5. For non-thinking: temperature=0.7, top_p=0.8, top_k=20, presence_penalty=1.5.
We're benchmarking and onboarding Qwen3.6-35B-A3B Uncensored Aggressive as a hosted, OpenAI-compatible API. Sign in for free credit and be ready when it lands, or tell us you want it and we'll prioritize it.