Unsloth Dynamic 3.0 GGUF Released

Unsloth released Dynamic v3.0, the latest version of its Dynamic quantization technology. The update achieved more than 10% higher accuracy for the Qwen3.8-27B model at the same size, representing a significant performance improvement over Dynamic v2.0, and is compatible with major inference engines such as llama.cpp and Unsloth Desktop.

The new methodology utilizes a high-quality imatrix calibration dataset collected from diverse sources. This dataset was refined for agentic coding, chat, and multilingual performance, with improvements to layer selection and quantization techniques. Notably, it was processed solely via Post-training quantization without QAT or QAD, preventing overfitting issues.

Key improvements include:

• The UD-Q2_K_XL model demonstrates 8% higher accuracy than the next highest model, reducing errors in HTML program generation.
• The UD-IQ1_S 1-bit quantized model retains approximately 72% accuracy despite being 89% smaller in size.
• MTP modules were removed from small quantized models under 8.37GB to save disk space.

For accuracy evaluation, Divergence-300 @32 was introduced, which better reflects actual inference paths than top-1% accuracy of single predictions. This method measures KL Divergence up to 32 tokens across 300 holdout examples to assess similarity to the BF16 model.