Qwen3.6-35B Hits 1M Downloads With a No-GPU Denoising Trick
A community-released GGUF blends an uncensored Qwen3.6 MoE base with Hermes function calling, using a math-based tensor repair method called Genesis.
Takeaways
- Community GGUF fusing uncensored Qwen3.6 35B MoE with Hermes function-calling finetune.
- Over 1 million downloads in first month, 579 likes on Hugging Face.
- Genesis method denoises tensors via Marchenko-Pastur SVD, no retraining required.
- 35B total, 3B active, 256 experts, 262K context extendable to 1M via YaRN.
- Hybrid Gated DeltaNet plus full attention at 3:1 ratio, 201 languages, multimodal.
- Runs on llama.cpp, LM Studio, Ollama, vLLM; recommended V8 APEX quant.
A 35B-parameter Qwen3.6 mixture-of-experts model has pulled over a million downloads in its first month on Hugging Face. Called Qwen3.6-35B-A3B-Uncensored-Genesis-Hermes-V11-GGUF, the release combines three components: an uncensored Qwen3.6 base from HauhauCS, a Hermes function-calling finetune from DJLougen, and a post-training numerical cleanup pass the author calls Genesis. It ships as a ready-to-run GGUF with vision support and agentic tool calling included.
What Genesis actually does
The premise behind Genesis is that training leaves random noise scattered through a model's tensors alongside real learned knowledge, and that noise makes outputs unstable and verbose. Rather than finetune the result, the author runs a numerical repair pass directly on the GGUF file in three stages.
First, it scans the state-space convolution tensors responsible for long-context memory and rebalances the heads. Second, it scans blocks in chunks and replaces zero-valued segments in broken tensors with better-matching ones, leaving the learned structure untouched. Third, it detects noise using a custom SVD calibrated to the Marchenko-Pastur distribution, a result from random matrix theory that predicts the eigenvalue spectrum of pure noise. Anything outside that band is treated as real signal and kept. Token embeddings, output weights, 1D tensors, biases, and norms are excluded from the SVD pass entirely. The author claims 99% of the signal and learned gradient survives.