Unsloth Dynamic 3.0 GGUFs

Unsloth has released its Dynamic v3.0 quantization technology and started supporting the Qwen3.8-27B model. It achieves over 10% higher top-1% accuracy compared to existing providers at the same capacity, and is compatible with major inference engines such as llama.cpp and Unsloth Desktop.

This update utilizes a high-quality imatrix calibration dataset collected from diversified sources to optimize agentic coding, chat, and multilingual performance. It is performed using pure PTQ (Post-Training Quantization) without QAT/QAD, minimizing overfitting issues, and the imatrix files have been made public for community testing.

Key technical improvements include:

• Introduction of Divergence-300 @32 metric: Evaluates similarity in actual inference paths by comparing 32-token sequences rather than single-token predictions.
• Removal of MTP module: Removing the MTP module from small quantized models under 8.37GB saves approximately 500MB of disk space.
• Support for ultra-small models: The 6.2GB UD-IQ1_S quantized model maintains approximately 72% top-1% accuracy while being reduced by 89% compared to the original.

It has recorded over 5.1 million downloads in 5 days and holds an advantage over competitors in KL Divergence benchmarks.