Unsloth launches Qwen3.8-27B with dynamic quantization v3, 10% higher top-1% accuracy

Unsloth released a new set of GGUF files for the Qwen3.8-27B model built on dynamic quantization v3.0. The company says the files achieve more than 10% better top-1% accuracy than any other provider at the same disk size, while remaining compatible with common inference engines such as llama.cpp and Unsloth Desktop. The updated methodology uses a high-quality calibration imatrix dataset drawn from diverse sources aimed at agent coding, chat, and multilingual tasks, together with improved layer selection and additional quantization techniques, all applied in a post-training step without QAT or QAD.

To avoid overfitting, Unsloth assembled an evaluation dataset of 300 held-out examples from Terminal-Bench 2.1, DeepSWE, Harbor, MathArena 2025-26 and long non-Latin instructions, none of which appear in the calibration dataset. The test runs a greedy argmax decoding for 32 tokens and compares BF16 output trajectories against all quantizers and providers using two metrics: average KLD (Kullback-Leibler divergence) and Divergence-300 @32. The graphs normalize disk-size comparisons by removing the MTP head from the axis to enable a fair comparison.

On the small-scale end, the UD-IQ1_S version weighs only 6.2 gigabyte (without MTP), a reduction of 89%, yet retains roughly 72% top-1% accuracy. The UD-Q2_K_XL version at 9.83 gigabyte shows an improvement of about 8% in top-1% accuracy over the next best alternative and succeeds in generating a functional HTML page with a single minor JS bug, whereas earlier versions failed completely. Removing the MTP module from quantizers below UD-Q2_K_XL saves roughly 500 megabyte, and the module is offered separately in the Q4_0 version for users who need it.

Unsloth reports 5.1 million downloads of the Qwen3.8 files within five days and releases the imatrix file for community research, evaluation, and fine-tuning. Compared internally to Dynamic v2.0, the KLD improvement is significant for the smaller quantizers but negligible for the larger ones, so the company continues to use v2 for large quantizers and plans further experiments. The entire process remains pure PTQ, with no training on the calibration dataset, reducing the risk of overfitting relative to QAD/QAT approaches.