text model · Phi-3.5 · iOS
Can I run Phi-3.5-mini 3.8B on iPad Pro M4 (16GB, 1TB/2TB config)?
Yes. Phi-3.5-mini 3.8B runs on iPad Pro M4 (16GB, 1TB/2TB config) at Q4_K_M (~3.7 GB of ~12 GB usable).
Runs at Q4_K_M using ~3.7 GB of ~12 GB usable. You have room for FP16 for higher quality.
That figure is at a 4k context and moves about ±15% as context length changes. iPad Pro M4 (16GB, 1TB/2TB config) leaves ~8.3 GB of headroom, room to step up to FP16 for higher quality.
- Q4_K_M needed
- ~3.7 GB
- Usable on device
- ~12 GB
- Device memory
- 16 GB
- Best quant
- Q4_K_M
Which quant fits
Running cost · estimate
- Power draw
- ~14 W
- Electricity / 1M tokens
- ~$0.02
- Pays for itself after
- ~3,331M tok
At ~$0.15/kWh and the estimated ~25 tok/s, a million generated tokens costs about $0.02 in electricity. Versus a hosted API at ~$0.5 per million tokens, the ~$1,599 iPad Pro M4 (16GB, 1TB/2TB config) pays for itself after roughly 3,331 million tokens, so local hardware is mostly a fixed cost, not a per-token one. TDP is the peak draw, so this is an upper bound. Assumptions.
How to run it
On iOS use Apple Foundation Models (Built into iOS 26, ~3B on-device model, zero download, fully private.).
- Parameters
- 3.82B
- Q4_K_M size
- 2.39 GB
- Q8_0 size
- 4.06 GB
- Context
- 128k
- Ollama tag
- phi3.5:3.8b
- Memory
- 16 GB unified
- Usable for weights
- ~12 GB
- Power draw
- ~14 W
- Best runtime
- MLX (via Python or Swift; mlx-lm package)
You could also run
Run Phi-3.5-mini 3.8B on other hardware
FAQ
Can iPad Pro M4 (16GB, 1TB/2TB config) run Phi-3.5-mini 3.8B?
Yes. Phi-3.5-mini 3.8B runs on iPad Pro M4 (16GB, 1TB/2TB config) at Q4_K_M (~3.7 GB of ~12 GB usable).
How much memory does Phi-3.5-mini 3.8B need?
iPad Pro M4 (16GB, 1TB/2TB config) has room to spare. At Q4_K_M the weights are ~2.39 GB; with KV cache and runtime overhead, budget ~3.7 GB at a 4k context.
What is the best tool to run Phi-3.5-mini 3.8B 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/phi-3.5-mini/ipad-pro-m4-16gb) Sources
- apple.com/ipad-pro
- apple.com/newsroom
- azure.microsoft.com
- developer.apple.com
- enclaveai.app
- github.com/a-ghorbani
- github.com/google-ai-edge
- github.com/mainframecomputer
- gsmarena.com
- huggingface.co
- layla-network.ai
- ollama.com
- phonearena.com
- privatellm.app
- support.apple.com/en-us/119891
- support.apple.com/en-us/119892
- techcommunity.microsoft.com
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.