OBSERVABILITY Signal 493
Unsloth Dynamic 3.0 GGUFs claim >10% better top-1% accuracy at same quant size
↗
Unsloth released Dynamic v3.0 GGUF quantization for Qwen3.8-27B, claiming over 10% better top-1% accuracy at the same model size compared to other providers, with improved KL Divergence and a new multi-token divergence metric.
Dynamic v3.0 lets teams run smaller quantized models that retain more of the original model's behavior, potentially reducing inference costs without sacrificing as much quality. The introduction of Divergence-300 @32 as a metric addresses a gap in evaluating whether quantized outputs actually follow the same trajectories as the full-precision model over multiple tokens, not just single-token accuracy.
Written by elseif from the cluster below · every claim links back to a sourceThe three things worth knowing
Dynamic v3.0 claims >10% better top-1% accuracy at the same disk size compared to other providers, with stronger KL Divergence results especially on smaller quants.
The MTP module is removed from quants under 8.37GB to save ~500MB, and a UD-IQ1_S quant at 6.2GB retains ~72% top-1% accuracy while being 89% smaller.
A new Divergence-300 @32 metric evaluates quantized output trajectories over 32 tokens against BF16 using 300 held-out examples, providing a measure of overfitting beyond single-token top-1% accuracy.
THE CLUSTER