Unsloth has released Dynamic v3.0 GGUFs, delivering over 10% better top-1% accuracy compared to other providers at the same model size. The new Qwen3.8-27B Dynamic v3.0 quants work with most inference engines including llama.cpp and Unsloth Desktop.

The methodology uses higher-quality imatrix calibration datasets refined for agentic coding, chat, and multilingual performance. Improvements include better layer selection and additional quantization techniques to preserve model quality. The team does not use QAT or QAD, relying purely on post-training quantization.

Notably, 5.1 million downloads were recorded in just 5 days. Smaller quants under UD-Q2_K_XL (8.37GB and lower) have removed the MTP module to save around 500MB of disk space. The UD-IQ1_S 1-bit quant is 6.2GB while retaining approximately 72% top-1% accuracy.

The team also introduced Divergence-300 @32 as a more effective metric using 300 held-out examples for greedy argmax decoding over 32 tokens. This measures KL Divergence to better gauge overfitting and output similarity to BF16 baselines. Results show UD-3 quants achieve up to +10% extra top-1% accuracy at all levels, especially on smaller quant sizes.