Qwen3.6-35B-A3B-Uncensored-Genesis-V2-APEX-MTP-GGUF Removes All Refusals
Qwen3.6-35B-A3B-Uncensored-Genesis-V2-APEX-MTP-GGUF is a fully quantized, refusal-free language model that packages the original Qwen3.6‑35B‑A3B MoE architecture into ready‑to‑run GGUF files. This release combines APEX and MTP‑APEX quantization formats with a numerical repair process that fixes silent drift and dead weight blocks inside the model’s recurrent layers. All the original multimodal capabilities remain intact — just the refusals have been removed and a critical long‑context bug on the output layers has been corrected.
LuffyTheFox created this version after diagnosing and surgically repairing three SSM layers that were too “loud” compared to their peers. By analyzing the raw tensor data with a method called Sig‑ScaleSync, the developer restored scale alignment and dramatically reduced distribution errors without any retraining. The result is a model that delivers the original reasoning performance but never says no to a user request.
Uncensored and repaired for long‑context use
Key Features
- Zero refusals across all tested prompts.
- Repaired dead blocks in long‑context memory layers.
- APEX quantization for fast consumer GPU inference.
- 35B parameters, only 3B active per token.
- 262K native context, extendable to 1M tokens.
- Native multimodal support (text, images, video).
This release is aimed at local AI users who need an unconditionally compliant model for creative writing, research, or private work. With the GGUF format, it runs smoothly on 12GB‑class cards like an RTX 3060, and the repaired SSM layers help keep long conversations or document analysis coherent. Anyone who values privacy and predictable output can run it entirely offline.
Developer insights and hardware notes
On a 12GB RTX 3060, the developer saw higher token rates without MTP, so the plain APEX quant is the practical speed choice for similar setups. The repair touched only 3 of 500 weight tensors, and after applying scale corrections the Wasserstein‑1 drift dropped by roughly 80%, leaving the model “clinically healthy” by the author’s metrics. Future work to regenerate more dead blocks depends on community support, as the developer currently operates on limited compute resources.
“Fully functional - 100% of what the original authors intended, just without refusals and with the critical architecture bug fixed on output layers.” — Source: Hugging Face
Project Links
