SC117 Strips Refusals From Qwen3.8-Flash-Next

Developer SC117 has published a modified GGUF build of Qwen3.8-Flash-Next designed to remove learned refusal behaviors while preserving the underlying quantization structure. The release quickly gained traction, surpassing 900,000 downloads on Hugging Face. To accomplish the edit without destroying the model's custom low-bit optimization, SC117 altered 144 refusal-direction projection tensors across 48 layers. Crucially, 1,079 non-target tensors were kept byte for byte, while nine target tensors were re-encoded using their original learned GSQ scales.

Traditional abliteration modifies a 16-bit BF16 checkpoint and re-quantizes the entire model, which destroys specialized low-bit scales. The upstream ISTA-DASLab build relied on GSQ and RCO optimization, combining a Gumbel-Softmax relaxation to learn grid assignments alongside resource-constrained optimization to pick per-tensor formats within a mixture-of-experts architecture. By editing tensors directly rather than re-quantizing from scratch, SC117 preserved those optimized parameters.

The resulting model is offered in four quantization tiers ranging from Q2_0 at 38 GB to IQ3_S at 55 GB, making each build just 0.25 to 0.44 GB larger than the upstream release. Performance remains essentially identical to the original, achieving a decode speed of 104.4 tok/s compared to the upstream speed of 104.5 tok/s. The release maintains a 262K context window and can run on a 12 GB GPU using Strata.

For practitioners, the edited weights work unmodified across popular inference frameworks including llama.cpp, Ollama, LM Studio, and vLLM, with full support for vision modalities and speculative decoding. However, because the build completely strips safety guardrails without providing replacement filters, engineers deploying the model must implement their own external moderation layers.