Qwen 3.6 35B uncensored multimodal weights hit HuggingFace in GGUF and FP8
LuffyTheFox released uncensored Qwen 3.6 35B weights in GGUF and FP8 formats on HuggingFace, enabling local vision-language inference on consumer GPUs without safety filters.
LuffyTheFox released uncensored weights for Qwen 3.6 35B, a multimodal mixture-of-experts model, on HuggingFace on May 13, 2026. The GGUF quantized variant has pulled 17,952 downloads, making it one of the more popular uncensored multimodal checkpoints in the open-weight ecosystem. The 35-billion-parameter architecture uses a mixture-of-experts design, activating subsets of parameters per token to reduce inference cost while maintaining model capacity. Vision capabilities let the model process images alongside text prompts—a feature set previously locked behind closed APIs or safety-tuned checkpoints. The uncensored release removes content filters, allowing practitioners to fine-tune or prompt the model for use cases that commercial providers restrict.
One practitioner ran the Q4 GGUF checkpoint on dual RTX 5060 Ti 16GB cards via LM Studio at full context, achieving 90 tokens per second. A second uncensored fine-tune, jLc2's HauhauCS-Aggressive variant, appeared on May 14, 2026, with similar multimodal capabilities. LuffyTheFox's FP8 safetensors variant also went live the same day, offering an alternative precision target for users with FP8-capable hardware.