Phi is Microsoft's family of small language models, built to squeeze strong reasoning and instruction-following into a compact footprint. That design goal makes the Phi Mini class one of the more interesting options to run Phi locally on a phone or laptop: it is small enough to fit on capable consumer hardware, yet punchy for its size. This tutorial walks through which Phi variant to pick for your iPhone or Mac, the GGUF file sizes and RAM each needs, the setup steps, and why running Phi on-device keeps every prompt on your device. If you are cross-shopping compact models in general, our best small language models roundup puts Phi next to its peers, and our Llama vs Phi vs Gemma on iPhone comparison weighs the families head to head.
Want the short version? Jump to the summary table. Want Phi-class models pre-packaged for on-device use? PocketLLM is built to handle it on iPhone, coming soon, join the launch list.
On a recent Pro-tier iPhone (around 8 GB of RAM) or an 8 GB Mac, run a Phi Mini model (roughly 3.8B parameters) at Q4, which is about 2.2 GB on disk. On a 16 GB Mac, a Q8 build gives a quality bump. Phi Mini sits at the larger end of the phone-friendly range, so on an older or base-tier phone a lighter 1B to 3B model is smoother. Running any of these locally is fully private: prompts never leave your device.
PocketLLM is launching soon. Private, on-device AI, starting on iPhone and iPad with more platforms planned. No account, no tracking, no cloud. Join the launch list and be first in.
Join the launch listWhich Phi variant fits your device
Microsoft has shipped several Phi releases over time, and the exact version numbers, benchmark scores, and release dates shift with each one. Rather than chase a single number, think in terms of the Mini class, which has consistently landed around 3.8B parameters and is the size that matters for on-device use. Confirm the specific build on its model card before you download, since capabilities and licensing can differ between releases.
Phi Mini (about 3.8B) at Q4, the everyday pick
This is the variant most people should run on-device. At Q4 it is roughly 2.2 GB on disk and wants around 4 to 5 GB of free RAM once the KV cache, runtime, and operating system are accounted for. It fits comfortably on a recent Pro-tier iPhone (approximately 8 GB of RAM) and any 8 GB or larger Mac, with quality strong enough for drafting, summarizing, and everyday chat.
Phi Mini at Q8, the quality pick
On a 16 GB Mac, step up to a Q8 build, roughly 4 GB on disk, for a modest quality bump over Q4. It is overkill for a phone and best reserved for machines with memory to spare.
A lighter alternative when 3.8B is too heavy
Phi Mini sits at the upper edge of what a phone handles comfortably, so the comfortable phone-class range of 1B to 3B parameters is a better fit for older hardware. On a base-tier or older iPhone, a smaller model from another family runs more smoothly. Our guides to running Qwen locally and running Gemma locally cover lighter picks in that size class.
Why file size is not the whole RAM story
A common mistake is reading the GGUF file size as the memory the model needs. The file is only the weights. When the model runs, it also holds a KV cache that grows with how long your conversation gets, plus the runtime itself and whatever the operating system is already using. iOS additionally caps how much memory a single app may use, which is stricter than a desktop. So a 2.2 GB file wants meaningfully more than 2.2 GB of free RAM in practice. Treat the file size as a floor, and leave headroom. A rough sizing rule for Q4 weights is about 0.6 GB per 1B parameters, which is why a 3.8B Phi Mini lands near 2.2 GB, then you add the extras on top.
How to run Phi locally on-device
- Pick by RAM. Recent Pro-tier iPhone or 8 GB Mac, Phi Mini at Q4. 16 GB Mac, Phi Mini at Q8. Older or base-tier phone, choose a lighter 1B to 3B model instead.
- Grab the GGUF. Phi Mini is published in the GGUF format that on-device runtimes read; choose the Q4 build for your size. New to quant tags? See our GGUF guide.
- Pick a local runtime. On Mac, LM Studio or Ollama load and run the GGUF offline. On iPhone, use a llama.cpp-based app.
- Confirm it is offline. Turn on Airplane Mode. Phi should still answer, which proves inference is fully local.
- Or skip the manual steps. An app that bundles compatible models picks a fitting quant and downloads it in one tap, with no GGUF handling on your part.
The summary table
| Variant | Quant | File size | RAM to run | Best device | Source |
|---|---|---|---|---|---|
| Phi Mini (~3.8B) | Q4 | ~2.2 GB | ~4-5 GB | Recent iPhone Pro, 8 GB+ Mac | Model card |
| Phi Mini (~3.8B) | Q8 | ~4 GB | ~6-7 GB | 16 GB Mac | Estimate |
| Lighter 1B-3B model | Q4 | ~0.7-2 GB | ~2-4 GB | Older or base iPhone | Estimate |
File sizes follow the rough 0.6 GB per 1B rule for Q4 weights and are approximate; RAM figures are estimates that assume normal context lengths, and both vary by build. Always check the model card for the specific quant you download.
What Phi is good at on-device
Phi's whole design brief is strong reasoning per parameter, so for its size it tends to do well on structured tasks: following instructions, working through short logic problems, drafting, and summarizing. The trade-off is the usual one for compact models, a smaller store of world knowledge than a large cloud model, and version-to-version differences you should verify on the model card. For everyday on-device work, though, a Phi Mini at Q4 is genuinely useful and stays private by construction. For a broader field of on-device options, see our roundup of the best on-device LLM apps for iPhone.
The privacy angle: local Phi vs hosted Phi
Running Phi locally is categorically private. The model and all inference live on your device, so prompts never leave it: no account, no server, no logging. That is different from any hosted Phi or cloud chat service, which sends your text to a remote server governed by that provider's policy. By downloading the open-weight model and running it offline, you get Phi's quality without handing your prompts to a third party.
| What matters | Hosted Phi / cloud chat | Manual local Phi (GGUF) | PocketLLM |
|---|---|---|---|
| Where it runs | Remote server | Your device | Your device |
| Account needed | Usually yes | No | No |
| Telemetry on chats | Per provider policy | None (local) | Zero |
| Works in Airplane Mode | No | Yes | Yes |
| Setup effort | Sign up | Pick GGUF plus runtime | One-tap (planned) |
A purpose-built on-device app makes local the default. PocketLLM is designed to run compatible models entirely on-device on iPhone, using a llama.cpp path for GGUF models, with zero telemetry on conversations.
Want Phi-class quality on your iPhone, fully on-device? PocketLLM is built to package compatible models as one-tap downloads with zero telemetry. Coming soon, join the launch list.
Frequently asked questions
Which Phi model can I run on an iPhone?
On an iPhone, stick to a Phi Mini class model, which is roughly 3.8B parameters and about 2.2 GB on disk at Q4. That fits best on a recent Pro-tier iPhone with around 8 GB of RAM, because the model needs headroom for the KV cache, the runtime, and the operating system on top of the file size. On an older or base-tier phone with less memory, a lighter 1B to 3B model from another family is a smoother pick. Version specifics change across Phi releases, so confirm the exact build you download.
How do I run Phi locally on a Mac?
To run Phi locally on a Mac, download a GGUF quant of a Phi Mini model, then load it in a local runtime such as LM Studio or Ollama. Pick the Q4 build for a good size-to-quality balance on an 8 GB Mac, or a Q8 build on a 16 GB Mac for a quality bump. Once it loads, turn on Airplane Mode and send a prompt to confirm inference is fully local. No account or internet connection is required after the file is on disk.
How much RAM do I need to run Phi?
A Phi Mini model at Q4 is about 2.2 GB as a file, but plan for more RAM than that. The running model also needs room for the KV cache, which grows with context length, plus the runtime and the operating system. As a rough guide, budget around 4 to 5 GB of free memory for a Q4 Phi Mini, and more for longer contexts or a Q8 build. File size on disk is a floor, not the full memory requirement.
Is running Phi locally private?
Yes. When you run Phi locally, the model and all inference live on your device, so your prompts never leave it: no account, no server, no logging. That is different from any hosted Phi or cloud chat service, which sends your text to a remote server governed by that provider's policy. Downloading the open-weight Phi model and running it offline gives you its quality with full on-device privacy.
Is Phi good for on-device use?
Phi is a capable small-model family from Microsoft, designed to pack strong reasoning and instruction-following into a compact size, which suits on-device use. The Mini class sits at the larger end of what a phone comfortably handles, so it shines most on recent Pro-tier iPhones and on Macs, while lighter families are easier on older phones. It converts to GGUF and runs in llama.cpp-based runtimes, making it a practical choice for private local AI.