Abliterated Gemma 3 and Qwen 3.5 checkpoints debut on HuggingFace
Three new abliterated checkpoints from arzaan789 strip safety filters from Google's Gemma 3 4B and Alibaba's Qwen 3.5 models, all quantized to GGUF and ready for local deployment.
Three abliterated checkpoints dropped on HuggingFace on May 12, stripping safety filters from Google's Gemma 3 4B and Alibaba's Qwen 3.5 0.8B and 4B models. Creator arzaan789 quantized each to GGUF format for local inference. All three support image-text-to-text pipelines and conversational use.
The Gemma 3 4B uncensored checkpoint is based on Google's gemma-3-4b-it instruction-tuned model. The two Qwen releases target Alibaba's Qwen 3.5 0.8B and 4B base models. Each model page on HuggingFace lists "abliterated" and "uncensored" among its tags, signaling removal of alignment layers that enforce content restrictions.
What stands out
- 01Size range for different hardware. The 0.8B Qwen variant fits on older GPUs or even high-end phones, while the 4B models balance capability and resource use on consumer desktops. GGUF quantization means practitioners can load them in llama.cpp, Ollama, or LM Studio without additional conversion.
- 02Multimodal abliteration. All three checkpoints retain image-text-to-text capability after abliteration. Most abliterated releases target text-only models; these preserve vision encoders, letting users run uncensored prompts on image inputs.
- 03Zero adoption so far. Each model page shows zero downloads and zero likes. That's typical for brand-new uploads, but it also means no community vetting yet — no benchmark runs, no user reports on output quality or how thoroughly the safety tuning was removed.
- 04No model cards. The HuggingFace pages carry only auto-generated metadata. There's no README explaining the abliteration method, no example prompts, no license file. Practitioners will need to infer usage and limitations from the base-model documentation.