ERNIE-4.5-21B-A3B-PT: hardware requirements

ERNIE-4.5-21B-A3B-PT needs about 15 GB of memory at Q4 quantization — a 13 GB download plus room for context and overhead. That fits a machine with 32GB RAM.

Parameters
22B
Tier
Needs a good desktop
Licence
See model card
Released
2025-08-12

This is a mixture-of-experts model. Only a fraction of its 22B parameters are used for each token, so it runs faster than its size suggests — but all of it still has to fit in memory. The speed is that of a small model; the memory requirement is that of a large one. More on this →

Will it run on your machine?

Download sizes by quantization

Lower quantization means a smaller file and less memory, at some cost to quality. Q4_K_M is the usual starting point. What is this?

QuantDownloadRAM @ 4KRAM @ 32KSource
Q4_K_Mrecommended13 GB15 GB17 GBmeasured
Q5_K_M16 GB17 GB19 GBestimated
Q6_K18 GB20 GB22 GBmeasured
Q8_023 GB26 GB27 GBmeasured
BF1644 GB47 GB49 GBestimated

“Measured” sizes are read from published GGUF files (lmstudio-community/ERNIE-4.5-21B-A3B-PT-GGUF). “Estimated” sizes are derived from the parameter count and are typically within a few percent. The RAM columns add the context cache and runtime overhead to the download size.

Other sizes of ERNIE 4.5 21B A3B