Phi 4 Mini on iPhone and Mac: Microsoft's small model for math and logic
· By Modelbase Team
Phi-4-mini is a small open language model from Microsoft with 3.8 billion parameters, built to be strong at reasoning, especially math and logic, while staying small enough for limited hardware. Microsoft released it in February 2025 under the MIT license, with a 128K-token context and support for more than 20 languages (Microsoft). In 4-bit form it’s a 2.18 GB download, which runs on recent iPhones and any Apple silicon Mac in an app like Modelbase, offline.
Phi 4 Mini at a glance
| Phi-4-mini-instruct | |
|---|---|
| Maker | Microsoft |
| Parameters | 3.8 billion |
| Context length | 128K tokens |
| Vocabulary | about 200,000 tokens |
| Languages | 23, from Arabic to Ukrainian; trained mostly on English |
| Training data cutoff | June 2024 |
| Released | February 2025 |
| License | MIT |
| Download in Modelbase (4-bit) | 2.18 GB |
Source: Microsoft’s model card; download size from Modelbase’s catalog.
The languages Microsoft lists are Arabic, Chinese, Czech, Danish, Dutch, English, Finnish, French, German, Hebrew, Hungarian, Italian, Japanese, Korean, Norwegian, Polish, Portuguese, Russian, Spanish, Swedish, Thai, Turkish and Ukrainian. The card notes that it’s trained mostly on English, so other languages may work less well.
Phi-4 vs Phi-4-mini
They’re different sizes of the same family:
| Phi-4 | Phi-4-mini | |
|---|---|---|
| Parameters | 14 billion | 3.8 billion |
| Context | 16K tokens | 128K tokens |
| Released | December 2024 | February 2025 |
| Runs on an iPhone | No, too large | Yes, on recent models |
Sources: Phi-4, Phi-4-mini.
Phi-4 needs a computer with plenty of memory. Phi-4-mini is the one Microsoft built for memory- and compute-constrained environments, such as phones and laptops.
What Phi 4 Mini is good at
Microsoft aims Phi-4-mini at situations where memory or compute is limited, where speed matters, and where you need strong reasoning, particularly math and logic (Microsoft). Modelbase’s catalog sums it up as “strong at math, logic and following instructions”. In practice:
- Step-by-step problems: percentages, unit conversions, word problems, scheduling puzzles.
- Following precise instructions: “Rewrite this in exactly three bullet points under 15 words each.”
- Structured writing: outlines, checklists, tables, summaries in a fixed format.
- Explaining: a concept, a formula, a paragraph you don’t understand.
Its limits, in Microsoft’s own words
Microsoft’s model card is candid: the model is “fundamentally limited by its size for certain tasks”, and it “does not have the capacity to store too much factual knowledge”, so you may run into factual errors. Its suggestion is to augment it with a search engine, particularly with retrieval (RAG), so the model has the facts to work from (Microsoft).
Modelbase does exactly that in two ways: web search, which gives the model sources for questions about the world, and documents: attach a PDF or a Word file and the app finds the passages that answer your question. See how to chat with a PDF on your iPhone.
Phi 4 Mini doesn’t read images and has no separate thinking mode. For questions about photos, use a vision model like Qwen 3.5 or Gemma 4 (E2B); for long step-by-step reasoning, try a reasoning model (see what a reasoning model is).
Can Phi 4 Mini run on an iPhone?
Yes, on recent ones. Modelbase recommends Phi 4 Mini (3.8B) for iPhone 15 Pro and newer, and tags it High CPU Usage: it’s one of the larger models in the catalog, so it works the phone harder than a 1B model and replies more slowly. On iPads and Macs with Apple silicon it has more room. On a Mac, the app considers a model a comfortable fit when it needs at most half of the Mac’s memory.
How to run Phi 4 Mini on iPhone, iPad and Mac
Modelbase runs open models on iPhone and iPad (iOS 18 or later) and Macs with Apple silicon (macOS 15 or later). It’s free, with no account.
- Install Modelbase from the App Store.
- Open the catalog: tap the model’s name at the top of the chat, then Manage models > Local.
- Search for “Phi” and tap Download on Phi 4 Mini (3.8B). At 2.18 GB, Wi-Fi is the better choice.
- Select it in the model picker. From then on it works offline.
Tip: give it clear, structured instructions. Phi 4 Mini is at its best when you tell it the format you want (“answer in a table”, “list the steps”, “give the final number on its own line”).
Prompts to try with Phi 4 Mini
A few examples that play to its strengths:
- “A recipe serves 4 and needs 350 g of flour. How much for 7 people? Show the steps, then the answer on its own line.”
- “I have meetings at 9:00 (45 min), 10:30 (1 h) and 13:15 (30 min). List my free blocks between 8:00 and 15:00.”
- “Explain compound interest to a 12-year-old in four sentences, then give one example with numbers.”
- “Turn these notes into a checklist with headings: …”
- “Is this argument valid? If not, say which step fails: …”
Coding and long chats
Two more notes from Microsoft’s model card. Most of Phi 4’s code training was in Python with common packages, so Microsoft recommends checking any code that uses other packages or languages by hand. And like other models, it can become repetitive or inconsistent in very long chat sessions; starting a new chat for a new topic helps (Microsoft).
Phi 4 Mini vs other models of its size
| Model | Maker | Parameters | Images | Thinking | Download |
|---|---|---|---|---|---|
| Phi 4 Mini | Microsoft | 3.8B | No | No | 2.18 GB |
| Qwen 3.5 (4B) | Qwen | 4B | Yes | Optional | 3.06 GB |
| Qwen 3 Instruct 2507 (4B) | Qwen | 4B | No | No | 2.28 GB |
| Llama 3.2 (3B) | Meta | 3.2B | No | No | 1.82 GB |
| Gemma 3 Text (4B) | 3.9B | No | No | 2.6 GB |
Sizes are the downloads in Modelbase’s catalog. Phi 4 Mini is the smallest download among the 4B-class models here, and the one Microsoft built specifically around math and logic. Qwen 3.5 (4B) adds images and an optional thinking mode at a larger size. For more comparisons, see which model to run on your iPhone, Llama 3.2 on iPhone and Mac and what a small language model is. New to all this? Start with how to run an LLM locally on iPhone, iPad and Mac.
FAQ
What is Phi-4-mini?
A 3.8-billion-parameter open language model from Microsoft, released in February 2025 under the MIT license, built for strong reasoning, especially math and logic, on limited hardware. It has a 128K-token context.
Is Phi-4-mini good?
For its size, at reasoning, math and following instructions, according to Microsoft. It’s weaker at factual knowledge, so pair it with web search or your own documents for facts.
Can I run Phi 4 Mini on my iPhone?
Yes, on iPhone 15 Pro and newer, in Modelbase. It’s a 2.18 GB download and works offline.
What’s the difference between Phi-4 and Phi-4-mini?
Phi-4 has 14 billion parameters and a 16K context; Phi-4-mini has 3.8 billion and a 128K context and is small enough for phones.
Is Phi-4-mini open source?
Its weights are released under the MIT license, which allows commercial use.
Does Phi 4 Mini support images?
No, Phi-4-mini-instruct is a text model. Use a vision model such as Qwen 3.5 or Gemma 4 (E2B) for photos.
Try it on your iPhone, iPad or Mac
Modelbase is free on the App Store.