Qwen2.5-72B-Instruct

Qwen
72.7B parameters
Released in bf16
Measured Sep 26, 2026

Qwen/Qwen2.5-72B-Instruct on Hugging Face

	bf16	Glyd	Change, %
Weights in GPU memory, GBEvery weight restored bit for bitlog	145.4	97.8	−33	log
48 GB GPUs it takesRTX A6000log	4	3		log
Tokens per second, 1 sequencebf16 on four A6000s, Glyd on threelog	4.5	6.4	+42	log
MMLU, 1,000 questions, %log	81.9	81.8		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	No
80 GB	H100, A100 80 GB	No	No
96 GB	RTX PRO 6000, GH200	No	Fits
141 GB	H200	Fits	Fits

Tokens per second, RTX A6000

Higher is better. bf16 on four GPUs, Glyd on three.

Sequences a step	bf16	Glyd	Change, %
	Tokens/s
1	4.5	6.4	+42
8	35.0	49.0	+40
32	134.6	153.8	+14
64	254.3	218.1	−14

log

Run it

pip install "glyd[gpu]"

import glyd
model = glyd.from_pretrained("Qwen/Qwen2.5-72B-Instruct")

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.