Huihui Strips Safety Limits From Qwen3.8-Flash-Next's 180B Model
Huihui released an uncensored 180B version of Qwen3.8-Flash-Next using abliteration, with GGUF quants supporting a 262K context window.
Takeaways
- Huihui released Huihui-Qwen3.8-Flash-Next-abliterated, a 180B uncensored Qwen3.8-Flash-Next.
- Built via abliteration: the refusal direction is projected out of activations without retraining.
- GGUF quants available with a 262,144-token context and full image-text-to-text support.
- Partial-layer ablation preserves reasoning, MTP and the vision encoder.
- Jetson AGX Orin deployment fits the model in roughly 34.5 GiB of VRAM via Orinfer E8P.
- Model card restricts use to research and controlled environments, not production.
Huihui has released Huihui-Qwen3.8-Flash-Next-abliterated, a weight-edited version of Qwen/Qwen3.8-Flash-Next designed to suppress learned refusal behavior. The checkpoint retains the base architecture’s multimodal input, long context, reasoning traces, prefix caching, and multi-token prediction. Huihui credits a sponsor with funding the ablation work and has published the model weights for download.
Developers get a 180B sparse mixture-of-experts model that accepts text and images while responding to prompts the base model may reject. Removing refusal behavior also weakens one safety mechanism and can reduce performance, making the checkpoint most relevant to evaluation, controlled research, and tightly sandboxed applications.
How the weight edit works
Abliteration changes model weights directly and requires no additional fine-tuning. The method estimates an activation direction associated with refusal, then modifies selected weight matrices to reduce that direction’s influence on the residual stream, the shared hidden state passed between transformer layers.
- Run contrasting prompts through the model and collect internal activations.
- Estimate the direction that most strongly separates refusal responses from compliant responses.
- Project that direction out of selected weights.
- Save the modified weights as a new checkpoint.
The underlying remove-refusals-with-transformers project describes its implementation as a rough proof of concept that works without TransformerLens. Abliteration lowers the tendency to produce standardized safety responses, while generation still varies with system prompts, decoding settings, and residual behavior elsewhere in the network.