KQuant quantization claims up to 3.7x better KL divergence for Qwen
ML engineer Alish released KQuant, an open-source quantization method offering MLX 4-bit packs for Qwen models that it says beat public baselines on KL divergence, by up to 3.7x for Qwen3.8-Flash-Next at a smaller size. The author says the quants run 6-15% faster on an M5 Pro, with GGUF versions to follow.