Nace quants vs Unsloth Dynamic v3.1, has anyone compared them on actual tasks?
retained text
Two quant updates landed in the last day: the Nace weight-space quants claiming up to 2x better KL divergence, and Unsloth's Dynamic v3.1 for Qwen3.8-27B. I've been trying to work out how much KL really tells you compared to running task evals. Has anyone put them side by side on real work like coding or tool calls? Which quant level are you on, and did you see any difference on long outputs?
document fetched 2026-11-15 00:36 · canonical https://reddit.com/r/LocalLLaMA/comments/1xvz070/nace_quants_vs_unsloth_dynamic_v31_has_anyone