Unsloth's Dynamic 3.0 quants squeeze 10% more accuracy from Qwen3.8 at the same size

Unsloth has released Dynamic v3.0, a post-training quantization method for large language models, debuting with GGUF builds of Qwen3.8-27B. The headline claim is more than 10% better top-1% accuracy at the same file size versus competing quantization providers, achieved without any quantization-aware training or distillation. The gains come from a higher-quality imatrix calibration dataset drawn from diverse sources and tuned for agentic coding, chat, and multilingual use, plus smarter per-layer selection and additional quantization techniques. The quants run on common inference stacks like llama.cpp and Unsloth Desktop, and the calibration imatrix is published for others to test and build on.

The more interesting part is how Unsloth measures quality. Rather than leaning on top-1% accuracy — essentially a single-token argmax that says little about real generation — or perplexity, which can mask errors through cancellation, they introduce Divergence-300 @32: 300 held-out prompts from benchmarks like Terminal-Bench 2.1, DeepSWE, Harbor, and MathArena, decoded greedily for 32 tokens and compared against the full BF16 model’s trajectory. Combined with KL Divergence, this is meant to catch overfitting and confirm the compressed models actually track the original’s behavior over multiple tokens. Notably, the team keeps their older v2.0 method for larger quants, where v3.0 showed little benefit.

The practical results are aggressive size cuts: a 1-bit UD-IQ1_S build lands at 6.2GB — 89% smaller than full precision — while retaining roughly 72% top-1% accuracy, and a 9.83GB UD-Q2_K_XL runs about 8% ahead of the next-best competitor and can now produce a working HTML program where it previously failed. Smaller quants also drop the MTP module to save disk space. For anyone running capable models on constrained local hardware, this continues Unsloth’s push to make heavily compressed LLMs usable without the accuracy collapse that typically accompanies low-bit quantization.