Qwen3.8-27B

Qwen
27.8B parameters
Dense
Apache-2.0
Released in bf16
Measured Oct 3, 2026

	bf16	Glyd	Change, %
Weights in vLLM on an H100 SXM, GiBMeasured, vLLM 0.30.0, v0.28.0log	50.22	38.77	−23	log
Weights as released, GBEvery weight restored bit for bitlog	55.56	37.34	−33	log
GPU memory in use, MiBnvidia-smi, H100 PCIe; with Glyd under a 48 GB card's 49,140log	51,771	41,071	−21	log
GPU time a token, 1 sequence, msH100 PCIelog	40.21	34.73	−14	log
MMLU, 300 questions, %Answers as bf16's on 99.67%log	79.67	80.00		log
Perplexity, enwik8log	15.1946	15.1941		log

Which single GPU it fits

Worked out from the measured weights.

Memory	GPUs	bf16	Glyd
16 GB	RTX 4080, RTX 5080	No	No
24 GB	RTX 4090, RTX 3090, A10	No	No
32 GB	RTX 5090	No	No
48 GB	RTX A6000, L40S, RTX 6000 Ada	No	Fits
80 GB	H100, A100 80 GB	Fits	Fits
96 GB	RTX PRO 6000, GH200	Fits	Fits
141 GB	H200	Fits	Fits

GPU time a token, H100 PCIe

Lower is better. The 12-bit layout; q, k, v and gate, up merged, as vLLM runs them.

Sequences a step	bf16	Glyd	Change, %
	GPU time a token, ms
1	40.21	34.73	−14
8	50.60	51.42	+2
32	85.19	84.52	−1

log

Run it

pip install "glyd[gpu]"

import glyd
model = glyd.from_pretrained("Qwen/Qwen3.8-27B")

Linux and an NVIDIA GPU, Ampere or later: the weights are packed on the GPU as the model loads, every one bit for bit (getting started). vLLM serves Glyd's weights too, for the models checked there (serving with vLLM).

Related models

Also measured

Twenty-one more open models measured, with quality wherever bf16 fits one GPU.