Unsloth · 2026-11-11 · notable
Unsloth Dynamic v3.1 — Qwen3.8-27B GGUF quants gain 5-10% accuracy
In short
Unsloth updated its Qwen3.8-27B GGUFs to the Dynamic v3.1 preview recipe. Unsloth says it improves accuracy by 5-10% at the same file size on most benchmarks. Most 2-bit to 6-bit quants were replaced.
Unsloth · 2026-11-11 · 3 sources checked
Unsloth's Dynamic v3.1 recipe makes the same-size Qwen3.8-27B GGUF files more accurate.
What is it?
Unsloth Dynamic v3.1 is a preview quantization recipe now used for the Qwen3.8-27B GGUFs on Hugging Face. Unsloth says it improves accuracy by 5-10% at the same size on most benchmarks. Most quants from 2-bit to 6-bit were updated, while the 3-bit UD3 quants stayed because they matched UD3.1.
How does it work?
The update lowers KL divergence for most quants and domains compared with Dynamic v3. Unsloth tested the files on Terminal Bench 2.1, DeepSWE v1.1, LiveCodeBench v6, GPQA Diamond, MMLU-Pro and AIME25, using 256K context for agentic tests and 128K for the rest. On AIME25 it reports 93-96% average accuracy recovery over 14 trials.
Why does it matter?
Local users who already run a Qwen3.8-27B quant get better answers without a bigger file. The page lists 16-19 GB of memory for the 4-bit build, so a single high-end GPU is enough. An NVFP4 build is also published.
Who is it for?
people running open models locally
Try it
unsloth run --model unsloth/Qwen3.8-27B-GGUF:UD-Q4_K_XL --temp 1.0 --top-p 0.95 --top-k 20 --min-p 0.0