text model · ERNIE · iOS
Can I run ERNIE 4.5 21B-A3B on iPhone Air?
No. ERNIE 4.5 21B-A3B needs ~14.3 GB even at Q4_K_M, but iPhone Air only has ~8 GB usable.
Needs ~14.3 GB even at Q4_K_M, but only ~8 GB is usable.
That figure is at a 4k context and moves about ±15% as context length changes.
The gap is about 6.3 GB: ERNIE 4.5 21B-A3B needs roughly 14.3 GB at Q4_K_M and iPhone Air leaves only about 8 GB usable for a model. The lightest tracked hardware that runs ERNIE 4.5 21B-A3B is the Nvidia GeForce RTX 4060 Ti (16GB) at 16 GB. See ERNIE 4.5 21B-A3B on Nvidia GeForce RTX 4060 Ti (16GB).
- Q4_K_M needed
- ~14.3 GB
- Usable on device
- ~8 GB
- Device memory
- 12 GB
Which quant fits
How to run it
On iOS use Apple Foundation Models (Built into iOS 26, ~3B on-device model, zero download, fully private.).
- Parameters
- 21B (MoE, 3B active)
- Q4_K_M size
- 12.42 GB
- Q8_0 size
- 21.61 GB
- Context
- 131k
- Memory
- 12 GB unified
- Usable for weights
- ~8 GB
- Best runtime
- llama.cpp + Metal (via PocketPal or Off Grid app)
What you can run instead
Run ERNIE 4.5 21B-A3B on other hardware
FAQ
Can iPhone Air run ERNIE 4.5 21B-A3B?
No. ERNIE 4.5 21B-A3B needs ~14.3 GB even at Q4_K_M, but iPhone Air only has ~8 GB usable.
How much memory does ERNIE 4.5 21B-A3B need?
iPhone Air does not have enough memory. At Q4_K_M the weights are ~12.42 GB; with KV cache and runtime overhead, budget ~14.3 GB at a 4k context. It is a Mixture-of-Experts model (21B total / 3B active), so all experts must stay in memory; memory tracks total params, not active params.
What is the best tool to run ERNIE 4.5 21B-A3B on iOS?
On iPhone and iPad, Apple Foundation Models (Built into iOS 26, ~3B on-device model, zero download, fully private.) is the standard choice. Phones realistically run 1B-4B class models. Anything larger thermally throttles or OOMs.
Embed this
A live badge for your model card or README, updated as the data is. [](https://localmodel.run/can-i-run/ernie-4.5-21b-a3b/iphone-air) Sources
Weights are measured from GGUF files; KV cache and overhead are computed, so totals can vary ~15% with context and runtime. Any tok/s is a bandwidth estimate. See methodology.