Sunday, 15 November 2026, story 5 of 7
Unsloth updates Qwen3.8-27B GGUFs with Dynamic V3.1 quantization
Unsloth says the preview algorithm adds 5-10% accuracy at the same size, in most 2-bit to 6-bit builds but not the 3-bit ones.

What happened
Unsloth has re-released its GGUF quantizations of Qwen3.8-27B using Dynamic V3.1, a preview version of its quantization algorithm, according to a page dated November 11. The company says average accuracy improves by 5-10% at comparable sizes, citing benchmarks including Terminal Bench 2.1, DeepSWE v1.1, LiveCodeBench v6 and GPQA Diamond. Most 2-bit to 6-bit builds changed. The 3-bit UD3 builds did not, because Unsloth found them on par.
The numbers are Unsloth's own. Its 4-bit builds need 16-19 GB of combined memory, plus 1-2 GB for MTP. The files are on Hugging Face and ModelScope. The source gives no price or license terms.
What it means
This is a vendor reporting its own benchmarks, run on its own methodology. The source does not say whether 5-10% means points or relative gain, and the algorithm is labelled a preview. Unsloth also says a better version is in progress, so these builds may be replaced again.
The agentic results came from 256K context and xhigh reasoning with preserved thinking. That is a heavier setup than many local runs, so gains may be smaller on yours. Unsloth also disputes a rival's quant benchmarks, saying that group likely ran one trial against its own 14. Only Unsloth's side appears in the source, but the lesson holds: single-run quant comparisons are noisy, so rerun them on your own tasks.
What to do: If you run a Qwen3.8-27B GGUF locally, re-download a 2-bit to 6-bit build and test it on your own tasks before replacing a working setup.
Source
Filed under Open source. Published .
"What it means" is analysis, not reporting. Found an error? Write to daily@vibitely.com.