Unsloth's Qwen3.8-27B Dynamic v3 GGUFs boost accuracy 10%

Unsloth has shipped updated Qwen3.8-27B GGUF quantizations built on a new Dynamic V3.0 quantization pipeline. According to the release, these quants deliver 10% higher accuracy at the same model size compared to previous versions, and outperform other quantization methods by more than 10% on the Div-300 and KLD benchmark suites. The update also introduces 1-bit quants that retain 77% of the original model's accuracy, enabling the 27B-parameter model to run on machines with just 8GB of RAM—a significant step for local deployment on modest hardware.

Notably, Unsloth clarifies that this is a pure improvement release, not a bug fix, and dismisses rumors of issues as fabricated. The team emphasizes that no training on the imatrix calibration dataset occurs, and neither QAT (quantization-aware training) nor QAD (quantization-aware distillation) is used—everything is achieved via post-training quantization. The imatrix file is publicly available for researchers and developers to test, evaluate, and create fine-tuned variations of Qwen3.8 using Unsloth's quants. Additionally, Unsloth announced an upcoming Desktop update featuring auto compaction and support for external APIs to handle tool calling, expanding the ecosystem around their optimized models.