Qwen3.5-122B abliterated weights debut on HuggingFace
Two independent abliterated versions of Alibaba's 122B-parameter mixture-of-experts model appeared on HuggingFace this week, removing safety filters from the multimodal architecture.
Qwen3.5-122B-A10B, Alibaba's 122-billion-parameter mixture-of-experts model with image-text-to-text capability, now has abliterated weights on HuggingFace. RobinsonLabs posted the first version on June 20, followed by morikomorizz's "Abliterix-Ultimate-Uncensored" variant on June 21. Both strip the safety tuning from the base model, which activates only 10 billion parameters per token despite the 122B total count. Abliteration rewrites the model's residual stream to remove refusal behavior without retraining, preserving underlying capability while eliminating safety-gated responses. The RobinsonLabs checkpoint ships in safetensors format for direct loading in transformers; the morikomorizz release includes GGUF quantized versions for llama.cpp and Ollama, making the model accessible to consumer GPU and CPU inference.
Qwen3.5-122B-A10B was released by Alibaba earlier this month as the largest open-weight multimodal model in the Qwen lineup, with the 10B active-parameter MoE design placing it in the same inference-cost bracket as dense 8B models. Both HuggingFace uploads are tagged for conversational and agent use cases. At 122B parameters, this is now the largest abliterated open-weight model available, exceeding the 70B ceiling of most abliterated Llama checkpoints.
