yeasah/Qwen3.8-27B-exl3 quants - lowest KLD of 0.0002
Summary
A new EXL3 quant of Qwen3.8-27B by yeasah achieves the lowest KLD (0.0002) among competitors while requiring only 18.8 GiB of VRAM for weights, outperforming quants from Unsloth and turboderp.
Similar Articles
Qwen 3.6 27B 30GB Same top p: 98.358 ± 0.033 % vs UD Q8 K XL 33GB Same top p: 97.426 ± 0.041 %
A community researcher shares a custom quantization recipe for Qwen3.6-27B that produces a smaller 30GB Q8 GGUF by keeping high-outlier sublayers in BF16, achieving better KLD and top-p metrics than Unsloth's 33GB Q8_K_XL variant.
I tested Qwen3.8 27B KLD for 21 Models at 9 different KV quants that fit on 16GB VRAM (189 benchmarks).
A community benchmark evaluated 21 Qwen3.8 27B model/KV quantization combinations (189 benchmarks) that fit in 16GB VRAM, finding q5_1/q5_1 is usually optimal and that upping model quant at q4_0/q4_0 beats lower quants at full f16 KV cache.
Qwen3.6-27B Quantization Benchmark
This article benchmarks various Qwen3.6-27B quantizations (Q8 to Q2) using KLD and Same Top P metrics, comparing providers like Unsloth and mradermacher, and offers recommendations for quality-size trade-offs.
Qwen3.8 Flash AP Quants
The post announces new quantized versions of the Qwen3.8 Flash model that outperform other high-quality quants through modified KLD measurement and optimized prefill performance.
Statistically Lossless Qwen3.8 27B GGUFs
kaitchup released statistically-lossless SLQ-quantized GGUF versions of Qwen3.8-27B in three accuracy-target variants (EAR98, EAR98.5, EAR99), built by mixing donor quantized weights across 498 matrices.