Qwen3.5-9B uncensored captioning and chat weights land on HuggingFace
Claymorecrystal released GGUF quantized weights for an unrestricted Qwen3.5-9B image-captioning fine-tune and a separate uncensored conversational variant, both flagged for local inference without content filters.
Two unrestricted Qwen3.5-9B checkpoints landed on HuggingFace this week from claymorecrystal, targeting practitioners running local inference without safety enforcement.
The first, qwen3.5-9b-nsfw-captioning-v5-GGUF, is a GGUF-quantized version of an uncensored image-captioning fine-tune based on oldhag88's work. GGUF quantization compresses the 9-billion-parameter model into CPU-friendly formats for local inference on consumer hardware. The checkpoint pulled 12 downloads since posting on June 20.
The second, Qwen3.5-9B-Uncensored-HauhauCS-Aggressive, is a separate unrestricted conversational variant built directly from Qwen/Qwen3.5-9B. The model card tags it as multilingual (English and Chinese) and explicitly uncensored. Both releases carry the "not-for-all-audiences" flag on HuggingFace, signaling that they lack the safety tuning present in the official Qwen3.5 release.
Qwen3.5-9B, Alibaba Cloud's 9-billion-parameter base model, shipped in late 2025 with strong multilingual performance and a 32,768-token context window. The official release includes safety alignment; these community fine-tunes remove that layer. Practitioners running local ComfyUI workflows or command-line inference can now caption images or run unrestricted chat sessions without server-side filtering.
The GGUF format has become the standard for CPU inference in the open-weight community. Quantized checkpoints trade precision for speed and memory efficiency, letting users run billion-parameter models on laptops and workstations without dedicated GPUs. The captioning fine-tune sits in a lineage that includes Florence-2, BLIP-2, and earlier Qwen-VL variants—all of which have seen uncensored community forks for describing visual content without content moderation.
Claymorecrystal's releases join a growing catalog of abliterated and unrestricted Qwen checkpoints on HuggingFace. The platform's "not-for-all-audiences" filter now surfaces hundreds of uncensored models across text, vision, and multimodal categories, reflecting steady demand from researchers and hobbyists who need unfiltered outputs for dataset labeling, creative workflows, and red-team testing.
