SYLIN

Unsloth Dynamic GGUF quants

by Unsloth · Technique · Text & code

promising

Small, mostly maker-reported gains over earlier quants; community finds quality about the same as v3.0

Dynamic v3.1 GGUF quants released for Qwen3.8-27B, with documented benchmark methods, graders and infra fixes

Why we say so

Unsloth claims 3.1 is 5-10% more accurate than UD3, GSQ-RCO and ExLlama-3 on Qwen3.8-27B. Commenters on Reddit and HN found the charts hard to read, and said the ExLlama-3 comparison depends on counting BF16 embeddings. One HN home benchmark found UD-Q2_K_XL 3.1 roughly tied with 3.0 (slightly worse at debugging, fits a 16 GB GPU), and another saw the same quality on a coding task.

Versions

VersionReleasedRank
Unsloth Dynamic GGUF quants 3.1LATEST
promising Claimed accuracy gains are unconfirmed; early user tests show parity with 3.0
Nov 12
Unsloth Dynamic GGUF quants 3.0
unproven Earlier release; the evidence here only compares it with 3.1
Aug 19

Evidence

  1. Nov 14
    Dynamic v3.1 GGUFs for Qwen3.8-27B claim 5–10% higher accuracy than UD3, GSQ-RCO and ExLlama-3; Unsloth disputes rival comparisons
  2. Nov 14
    Commenter found the chart unclear for choosing a quant on 24–32GB GPUs; Unsloth says main plots exclude embeddings, which ExLlama-3 keeps in BF16
  3. Aug 28
    Unsloth dynamic 3-bit GLM 5.3 Flash quant stayed coherent in reasoning and code, with no looping or hallucination
  4. Aug 19
    Dynamic 3.0 GGUFs released for Qwen3.8-27B, using an updated imatrix dataset and layer selection, no QAT or training on calibration data
  5. Aug 19
    Claims over 10% better top-1 accuracy than other providers' quants of Qwen3.8-27B at the same size

Cost to run

Alternatives

QwenQwen3.8-27B UD-Q2_K_XL v3.1 roughly matches v3.0 on a home benchmark: slightly worse at debugging, 46.5k vs 61.7k tokens per task, fits 16 GB GPU