Huihui Strips Alibaba's 177B Qwen3.8-Flash-Next of All Refusals
A 177B-parameter Qwen3.8-Flash-Next variant ships in GGUF form with refusal behavior surgically removed, ready to run locally through llama.cpp.
- Huihui-ai released an abliterated GGUF of Qwen3.8-Flash-Next, a 177B-parameter MoE model.
- Abliteration projects the refusal direction out of model weights via orthogonalization, no retraining required.
- Runs locally on llama.cpp, Ollama, LM Studio, vLLM, and SGLang with OpenAI-compatible APIs.
- UD-Q4_K_XL quantization is 111 GB with a 262,144-token context window.
- Typical quality degradation from abliteration is roughly 1-3% versus the base model.
- Free under qwen-community-1.0 license; already at 166K+ downloads.
The Huihui-ai profile has published a weight-edited version of Alibaba’s Qwen3.8-Flash-Next model. The GGUF repository packages the model for local inference through llama.cpp and compatible applications, with reduced refusal behavior as its defining modification. The repository displayed roughly 166,000 downloads at publication, although Hugging Face download counts represent qualifying file requests rather than unique users or completed installations.
| Detail | Repository claim |
|---|---|
| Base model | Qwen3.8-Flash-Next |
| Architecture | 177B-parameter qwen4exp mixture of experts |
| Modification | Refusal-direction abliteration |
| Distribution format | GGUF quantizations |
| Featured quantization | UD-Q4_K_XL, approximately 111 GB |
| Advertised context | Up to 262,144 tokens |
| License | qwen-community-1.0 |
How one vector becomes a weight edit
Abliteration uses an empirical finding from the original study: refusal behavior in several instruction-tuned models correlates with a direction in the residual stream, the shared hidden state passed between transformer layers. Removing that direction can suppress many explicit refusals while retaining much of the model’s previous behavior.
The procedure generally follows four steps:
- Run matched sets of refusal-triggering and benign prompts through the model.
- Record hidden activations at selected layers and token positions.
- Calculate candidate refusal directions from the difference between the two activation sets.
- Project the selected direction out of matrices that write to the residual stream, including embeddings, attention output projections, and MLP output projections.
With a unit refusal vector r_hat, the conceptual projection is:
P = I - r_hat @ r_hat.T
h_projected = P @ hThe implementation applies the equivalent projection to each targeted weight matrix. The multiplication side depends on how that matrix stores its output dimension. This edit suppresses the measured linear feature; refusal behavior can also depend on other directions, prompt distributions, decoding settings, and post-training biases.
The Huihui build cites Sumandora’s
This story is for Pro members
You've reached the end of the free preview. Upgrade to AlphaSignal Pro to read the full article - and everything else behind the paywall.
Comments
