Glyd
GitHubGet started

Llama-4-Scout-17B-16E-Instruct

  • Meta
  • 108.6B parameters
  • Mixture of experts, 17B active
  • Released in bf16
  • Measured Sep 27, 2026

meta-llama/Llama-4-Scout-17B-16E-Instruct on Hugging Face

Raw logsRun it

Llama-4-Scout-17B-16E-Instruct with Glyd

bf16GlydChange, %
Weights in GPU memory, GBEvery weight restored bit for bitlog217.28145.80−33log
Smallest single GPU, GBWeights, an 8K-token KV cache and 1.5 GB for the runtime—141
H100 80 GB GPUs it takes32
Its matrices, GBPacked and unpacked, every one bit for bitlog211.9142.2−33log

Which single GPU it fits

Worked out from the measured weights.

MemoryGPUsbf16Glyd
16 GBRTX 4080, RTX 5080NoNo
24 GBRTX 4090, RTX 3090, A10NoNo
32 GBRTX 5090NoNo
48 GBRTX A6000, L40S, RTX 6000 AdaNoNo
80 GBH100, A100 80 GBNoNo
96 GBRTX PRO 6000, GH200NoNo
141 GBH200NoFits

Run it

pip install "glyd[gpu]"
import glyd
model = glyd.from_pretrained("meta-llama/Llama-4-Scout-17B-16E-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). Its experts are packed too (v0.22.0). 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.