A 144-tensor refusal edit avoids full requantization

SC117 has published an abliterated Qwen3.8-Flash-Next GGUF designed to reduce learned refusal behavior while preserving the upstream model’s quantized values outside the edited tensor set. The release changes 144 projection tensors across 48 layers, retains 1,079 non-target tensors byte for byte, and re-encodes nine targets with their learned GSQ scales.

At publication, the SC117 repository had passed 900,000 downloads and was near the top of Hugging Face’s trending list. Its technical contribution is a way to apply a narrow post-training edit to a model whose learned quantization cannot be reproduced by a conventional BF16-to-GGUF conversion.

Why full requantization loses the original

Abliteration is the informal name for a low-rank weight edit that projects an activation direction associated with refusals out of selected model matrices. A rank-1 edit uses one vector direction, making the correction much smaller and more structured than a general weight update.

A conventional workflow applies the edit to BF16, a 16-bit floating-point checkpoint, and then quantizes the entire model again. That process cannot preserve the exact low-bit values in this release because the upstream quantization was learned during optimization.

The upstream ISTA-DASLab build combines GSQ and RCO. GSQ learns grid assignments and per-group scales for each tensor through a Gumbel-Softmax relaxation. RCO then selects a quantization type for each tensor while keeping the complete model within a size budget. The model uses a mixture-of-experts architecture, which routes tokens through selected expert feed-forward blocks instead of activating every expert for every token.

Quantized values produced by GSQ-RCO depend on the learned codes, scales, and tensor-level format choices. Running a standard quantizer over BF16 weights would produce a different low-bit checkpoint and discard those optimized parameters.