Skip to main content
Back

LuffyTheFox's Qwen3.6 Hits 1M Downloads With Zero-Refusal Uncensored Local AI

A community-built GGUF stacks an uncensored Qwen 35B MoE with a Hermes agent transfer and a tensor-level noise repair pass called Genesis.

Model, Full Access
  • V10 GGUF release of an uncensored Qwen 35B A3B MoE with Hermes agent transfer.
  • 35B total, 3B active, 256 experts, 262K context extendable to 1M with YaRN.
  • Genesis method: SVD-based tensor repair on raw GGUF bytes, no retraining required.
  • Base model HauhauCS Aggressive reports 0 refusals out of 465 prompts.
  • Ships with mmproj for vision; runs in llama.cpp, Ollama, LM Studio, vLLM under Apache 2.0.
  • Over 1.1M downloads on Hugging Face; author recommends 27B Genesis for coding.

A community merge on Hugging Face has quietly crossed the million-download mark. Qwen3.6-35B-A3B-Uncensored-Genesis-Hermes V10 is a GGUF build that layers three things on top of a Qwen 35B mixture-of-experts base: an uncensored fine-tune, a Hermes function-calling transfer, and a post-training tensor repair pass the author calls Genesis. It ships under Apache 2.0 and runs in every mainstream local runtime.

What is actually in the file

The model is a Qwen 35B MoE with roughly 3B active parameters per forward pass and 256 experts, of which 8 are routed and 1 is shared per token. The architecture mixes Gated DeltaNet linear attention with full softmax attention at a 3:1 ratio across 40 layers. Native context is 262K tokens, extendable to 1M with YaRN, and the model handles text, image, and video natively with a 248K vocabulary covering 201 languages.

The starting point is HauhauCS's aggressive uncensored variant, which reports 0 refusals out of 465 test prompts. On top of that base, the author transferred roughly 2k blocks from two FFN expert tensors of a Hermes-finetuned Qwen 3.5 GGUF to graft agentic behavior onto an otherwise non-agentic model.