Nace releases weight-space quantized GGUF weights for Qwen models
Nace released GGUF weights for Qwen3.5-9B, Qwen3.8-27B, Qwen3.6-35B-A3B and Qwen3.8-Flash-Next, produced through a quantization process that uses AI agents to propose and evaluate tensor-precision changes. Nace claims its files deliver up to twice better KL divergence than alternatives, but these results are vendor-reported; its MMLU check showed gains that were mostly within noise. A separate Reddit discussion asked whether the quantizations outperform Unsloth Dynamic v3.1 on real coding, tool-call and long-output workloads, with no comparison results provided.
The release offers a new weight-space-programming approach to producing GGUF quantizations, but its claimed advantages over existing quants still require independent task-level validation.
#Nace#Kapitsa#Qwen3.5-9B#Qwen3.8-27B#Qwen3.6-35B-A3B#Qwen3.8-Flash-Next#Nace WSP#GGUF#Unsloth Dynamic v3.1