Hacker News
Unsloth Dynamic 3.0 GGUFs Ship a 6.2GB 1-Bit Quant, and the HN Thread Pins Them on Publishing KL Divergence Instead of Benchmarks
Unsloth released Dynamic 3.0 quantization on 2026-08-19 with new imatrix calibration sets tuned for agentic coding and multilingual work, claiming up to 10% better top-1 accuracy at equal file size versus other providers, roughly 8% better at 9.83GB for Qwen3.8-27B Q2_K_XL, and a new UD-IQ1_S 1-bit quant at 6.2GB holding about 72% top-1. In 100 HN comments the sharpest pushback came from lostmsu: 'KLD isn't how that works either. The truth is in the middle and they aren't showing it,' with others noting no real-task benchmarks were published. Daniel Hanchen replied that MTP was only stripped from IQ2_XXS and below to save 500-750MB, and that they ran a 'Divergence-300 @32' eval with larger suites planned.
↳ Follow the thread