OpenModelsHQ

ERNIE-4.5-21B-A3B-PT

Baidu ERNIELLMMoEApache-2.0China Desk

ERNIE-4.5-21B-A3B-PT is an open-weight language model from Baidu ERNIE with 21.9B parameters (mixture of experts). At 4-bit it needs about 15 GB of memory, so the smallest current device that fits it is the Arc Pro B60.

Overview

Estimated memory at 4-bit
15GB

Includes room for an 8k context and runtime overhead. More context needs more memory.

Total parameters
21.9B
Mixture of experts, all weights counted
Context window
128K
131,072 tokens
Cheapest API, blended / 1M
No hosted API
Weights only

Key information

Lab
Baidu ERNIE
Released
28 Jun 2025
Parameters
21.9B · MoE
Type
Language
Context
131,072 tokens
License
Apache-2.0
HF downloads, 30 days
16,524
View all

Will it run on your hardware?

How we estimate
Memory needed vs devices
ERNIE-4.5-21B-A3B-PT needs about 15 GB at 4-bit, 26.1 GB at 8-bit and 48.8 GB at 16-bit. At 4-bit it fits GeForce RTX 5090, DGX Spark, Mac Studio M5 Ultra (256GB). Dashed lines are usable memory on each device; a model fits when its estimate stays within 90% of it. Log scale. 8k context.
Memory needed vs devices: data
Estimate or deviceMemory
Q415 GB
Q826.1 GB
BF1648.8 GB
RTX 5060 Ti 16 GB16 GB usable
RTX 5090 32 GB32 GB usable
DGX Spark 128 GB108.8 GB usable
Mac Studio M5 Ultra 256 GB192 GB usable

Single-device memory estimate. Pick a device to update the On your device column.

36 GB usable memory

FormatMemory neededOn your deviceExample hardware
4-bit (Q4)≈ 15 GBFits 42%Arc Pro B6024 GB memory
8-bit (Q8)≈ 26.1 GBFits 73%Radeon AI PRO R970032 GB memory
16-bit (BF16)≈ 48.8 GBToo big 136%A100 80GB SXM80 GB memory

Fits means the estimate stays within 90% of usable memory. Tight is above 90% and up to 100%. Too big is above 100%. These are memory checks, not speed or compatibility tests.

Mac allowances are 66% of total RAM up to 36 GB, and 75% above. AI boxes allow 85% because their memory is shared with the operating system.

Where to run it

On a Mac

Start with the MLX files, built for Apple Silicon.

mlx-community/ERNIE-4.5-21B-A3B-PT-4bitCheck your Mac

With GGUF

Choose a file for llama.cpp or a compatible app.

unsloth/ERNIE-4.5-21B-A3B-PT-GGUFFind compatible hardware

Through an API

No hosted API found.

MLX on Apple SiliconView repository
mlx_lm.generate --model mlx-community/ERNIE-4.5-21B-A3B-PT-4bit \
  --prompt "Explain how a transformer works."
GGUF with llama.cppView repository
llama-cli -hf unsloth/ERNIE-4.5-21B-A3B-PT-GGUF \
  -c 8192 -p "Explain how a transformer works."

Example commands built from the repository ids above. They need a working install of MLX or llama.cpp. We have not run them; check each repository for available files.

More about this model

Frequently asked questions

How much memory does ERNIE-4.5-21B-A3B-PT need? At an 8k context, about 15 GB at 4-bit, 26.1 GB at 8-bit and 48.8 GB at 16-bit (BF16). A longer context needs more.

Can I run ERNIE-4.5-21B-A3B-PT on a Mac? The smallest current Mac that fits the 4-bit estimate of 15 GB is the Mac mini M6 (32GB). An MLX build is available: mlx-community/ERNIE-4.5-21B-A3B-PT-4bit.

Does the full context fit? ERNIE-4.5-21B-A3B-PT supports up to 131,072 tokens of context. Our memory estimates use 8k, so long prompts need more memory than shown.

What does it cost through an API? We did not find a hosted API for this model.

License in plain English

Apache-2.0. This license permits commercial use, modification and redistribution, with conditions such as keeping the required notices. This summary is not legal advice.

Read Apache-2.0

Read about this license

OMHQ RigBench

Not yet tested. We show no speed or quality score without a measurement.

Revision history and sources

Last updated . Sources:Hugging Face model card, MLX repository, GGUF repository.