Skip to main content
Back

LuffyTheFox's Qwen3.6 Merges Uncensored AI and Tool-Calling Into One Local Model

A community-built GGUF release fuses an uncensored Qwen3.6-35B MoE base with Hermes function-calling weights and a custom tensor-denoising pass.

Model, Full Access
  • Community GGUF release Qwen3.6-35B-A3B-Uncensored-Genesis-Hermes-Final-GGUF has crossed 923K downloads on Hugging Face.
  • Built on the uncensored HauhauCS Qwen3.6-35B-A3B base with 0/465 reported refusals.
  • Hermes function-calling behavior added by transplanting ~2k blocks from two FFN expert tensors.
  • Author's Genesis pass denoises tensors via Marchenko-Pastur SVD without retraining the model.
  • MoE architecture: 35B total, ~3B active, 256 experts, 262K native context, multimodal.
  • NVFP4 quant runs on a single 24 GB GPU via llama.cpp, Ollama, LM Studio, vLLM.

A community release on Hugging Face is combining three components that usually live in separate repos: an uncensored Qwen3.6-35B-A3B base, Hermes-style function-calling weights, and a post-training tensor cleanup pass the author calls Genesis. The package ships as Qwen3.6-35B-A3B-Uncensored-Genesis-Hermes-Final-GGUF and has pulled roughly 923K downloads, making it one of the more actively grabbed local-inference bundles on the Hub.

What's Actually in the Box

The release is a GGUF quantization built on HauhauCS's uncensored variant of Qwen3.6-35B-A3B, which the base author reports as 0/465 refusals. On top of that, roughly 2K blocks from two FFN expert tensors in a Hermes finetune were transplanted into the uncensored base, giving the model Hermes-style tool-calling behavior without a full retrain. The architecture underneath is standard Qwen3.6 MoE:

  • 35B total parameters, ~3B active per forward pass
  • 256 experts, 8 routed plus 1 shared per token
  • Hybrid Gated DeltaNet linear attention plus full softmax attention in a 3:1 ratio