Five uncensored Qwen3.6-35B fine-tunes surface on HuggingFace in 24 hours
At least five uncensored Qwen3.6-35B fine-tunes appeared on HuggingFace between June 27–28, including multimodal GGUF quants and a mixture-of-experts checkpoint from Dendrite Holdings.
Five uncensored fine-tunes of Qwen3.6-35B appeared on HuggingFace between June 27 and June 28, 2026. The releases include dendriteholdings/albedo-qwen3.6-35b-king-XXX, a mixture-of-experts checkpoint tagged "not-for-all-audiences," and four variants of the Qwen3.6-35B-A3B-Uncensored-HauhauCS-Aggressive base by different uploaders—wang-yang, sadii0101, ddgiiddg, and hh676. All five carried zero downloads and zero likes at publication, suggesting they went live within hours of each other.
The wang-yang upload ships as a GGUF quantization optimized for llama.cpp and speculative decoding. The sadii0101, ddgiiddg, and hh676 variants are tagged for image-text-to-text pipelines, indicating multimodal capability. All four HauhauCS-Aggressive variants explicitly label themselves "uncensored" and support English generation. The Dendrite Holdings albedo checkpoint uses the qwen3_5_moe architecture and carries an XXX suffix, though no model card details are visible yet.
Format and capability breakdown
Each upload targets a different use case. The wang-yang GGUF is built for CPU/GPU hybrid inference with llama.cpp's draft-model acceleration. The three multimodal variants add vision encoders, letting users prompt with images alongside text. The Dendrite albedo checkpoint is the only one explicitly flagged as MoE, though all five derive from Qwen3.6's 35B parameter base. None of the uploaders have published benchmark numbers, training recipes, or ablation details.
The timing and naming overlap suggest coordinated experimentation or a shared upstream fine-tune that multiple users quantized and re-uploaded. All five are safetensors or GGUF format, ready for local inference without API keys or rate limits.
