Short answer
Memory decides. A model needs roughly its file size in memory, plus room for the conversation, and a phone or tablet gives one app only part of its memory. As a guide: an iPhone with 8 GB (iPhone 15 Pro, every iPhone 16 and 17) runs models of about 4 billion parameters; a 12 GB iPhone (17 Pro, Air, 18 Pro) runs 8B to 9B; older 4 to 6 GB iPhones suit 1B to 2B.
On a Mac the same rule scales up: 8 GB runs 2B to 4B, 16 GB runs 9B, 24 to 36 GB runs 27B to 35B, and 96 GB or more runs models above 100B. Apps such as Noema show an estimated fit for your device before you download anything.
How much memory a model needs
A model's size on disk is a good first estimate of the memory it needs. Most people run models quantized to about 4 bits per weight (the file is labelled Q4_K_M for GGUF or 4-bit for MLX), which is roughly a quarter of the full-precision size with a small loss in quality.
| Model size | File at Q4_K_M | Example |
|---|---|---|
| Under 1B | About 0.5 GB | Qwen3.5-0.8B, 0.53 GB |
| 2B | About 1.3 GB | Qwen3.5-2B, 1.28 GB |
| 4B | 2.5 to 2.7 GB | Qwen3-4B, 2.50 GB; Qwen3.5-4B, 2.74 GB |
| 8B to 9B | 5.0 to 5.7 GB | Qwen3-8B, 5.03 GB; Qwen3.5-9B, 5.68 GB |
| 14B | About 9 GB | Qwen3-14B, 9.00 GB |
| 27B | About 17 GB | Qwen3.8-27B, 16.81 GB |
| 30B to 35B | 19 to 22 GB | Qwen3-30B-A3B, 18.56 GB; Qwen3.5-35B-A3B, 22.02 GB |
| 122B | About 74 GB | Qwen3.5-122B-A10B, 74.27 GB |
On top of the file, the model needs memory for its context: the conversation so far, attached documents and tool definitions. Apple's MLX team measured Qwen3 models at 4-bit with a 4,096-token prompt and found each used a little more than its file size: 5.61 GB for the 8B and 9.16 GB for the 14B. Longer contexts need more, and vision or audio support adds its own files.
iPhone
iOS does not let one app use all of a phone's memory, so a model has to be well under the installed amount. Apple does not publish iPhone memory; the figures below are the ones reported in Wikipedia's list of iPhone models. Noema runs on iOS 18, which covers every iPhone in this table.
| Memory | iPhones | Model size to plan for |
|---|---|---|
| 3 to 4 GB | XR, XS, 11 series, SE (2nd and 3rd generation), 12 and 12 mini, 13 and 13 mini | Up to about 1B to 2B |
| 6 GB | 12 Pro, 13 Pro, 14 and 14 Pro series, 15 and 15 Plus | 2B comfortably; a 4B is tight |
| 8 GB | 15 Pro and 15 Pro Max, 16 series, 16e, 17, 17e | 4B |
| 12 GB | 17 Pro and 17 Pro Max, Air, 18 Pro and 18 Pro Max, Duo | 8B to 9B |
A useful marker: the iPhones that support Apple Intelligence, from iPhone 15 Pro onward, are the ones with 8 GB or more. Noema's guided setup starts devices with 8 GB or more on Noema 1.5 4B Deep, a 4B model.
iPad and Apple Vision Pro
Apple lists memory for iPads with M-series chips and for Vision Pro. Figures for A-series iPads are the ones Wikipedia's list of iPad models reports.
| Memory | Devices | Model size to plan for |
|---|---|---|
| 3 to 4 GB | iPad 7th to 10th generation, iPad mini 5 and 6, iPad Air 3 and 4, iPad Pro (2018) | Up to about 1B to 2B |
| 6 GB | iPad (A16), iPad Pro (2020) | 2B comfortably; a 4B is tight |
| 8 GB | iPad mini (A17 Pro); iPad Air with M1, M2 or M3; iPad Pro with M1, M2 or M4 and less than 1 TB of storage | 4B |
| 12 GB | iPad Air (M4); iPad Pro (M5) with less than 1 TB of storage; iPad mini (A20 Pro), as reported | 8B to 9B |
| 16 GB | iPad Pro with M1, M2, M4 or M5 and 1 TB or 2 TB of storage; Apple Vision Pro (M2 and M5) | 8B to 9B with a long context |
Mac
A Mac shares its memory between macOS, your apps and the model, and it can use a much larger share of it for a model than a phone can. These are the Macs Apple sells today; an older Apple silicon Mac with the same memory behaves similarly.
| Memory | Current Macs | Model size to plan for |
|---|---|---|
| 8 GB | MacBook Neo | 2B comfortably; a 4B with a short context |
| 16 GB | MacBook Air (M5), 14-inch MacBook Pro (M6), iMac (M6) and Mac mini (M6), base configurations | 9B; a 14B at 4-bit is tight |
| 24 GB | Upgrades of the above; MacBook Pro and Mac mini with M5 Pro, base | 14B easily; a 27B with a modest context |
| 32 to 36 GB | Top upgrade of the M5 and M6 models; MacBook Pro and Mac Studio with M5 Max, base | 27B to 35B |
| 48 to 64 GB | M5 Pro and M5 Max upgrades | 27B to 35B at 8-bit, or with very long contexts |
| 96 GB and more | M5 Max at 128 GB; Mac Studio with M5 Ultra, 96 to 512 GB | Models above 100B, such as Qwen3.5-122B-A10B at 4-bit |
How to run Qwen models locally on a Mac turns this table into commands for one model family.
What each size is good for
- 1B to 2B: quick replies, rewriting, simple questions, and tasks where speed matters more than depth.
- 4B: an everyday assistant on a phone: summaries, drafting, explanations and questions about documents you provide.
- 8B to 14B: noticeably stronger reasoning and coding help; the practical ceiling for phones and tablets.
- 27B and larger: the closest a personal machine gets to large cloud models, for Macs with 24 GB or more.
Treat that as a rough guide. Newer small models often beat older large ones, and a model tuned for a task can outperform a bigger general one.
Speeds we can point to
We do not publish a table of speeds per device. Speed depends on the chip, the model, its quantization, the context length, the runtime and even the phone's temperature, and a single number would mislead. Two measurements with a named source:
- An independent user reported roughly 23 tokens per second from an 8B model at Q4 on a 12 GB iPad with M5, using Noema (October 2025).
- In NoemaAI's Overfit field test, Gemma 4 26B-A4B, a model larger than the phone's memory, generated 3.5 tokens per second on an iPhone, with a 20-second wait for the first token.
To measure your own device, use the benchmark built into the app you run. Noema's benchmark warms up, measures three runs and reports the median speed and peak memory; on Mac it is in Tools › Benchmarking Center.
If the model you want does not fit
- Shorten the context first. A long context can use more memory than expected, even with a small model.
- Try a lighter quantization, or the next size down. A smaller model that fits is better than a larger one that swaps or fails to load.
- Use your Mac's memory from your phone. With Constellation, Noema on iPhone or iPad can use the models on your awake Mac. Other apps offer a similar link to a computer.
- Page the model from storage. Overfit runs compatible mixture-of-experts models whose full weights do not fit in memory by keeping the shared core in memory and reading experts from storage. It is slower, as the field test above shows.
Check before you download
You do not have to work this out by hand. In Noema, each model page in Explore shows compatibility and an estimated memory fit for the device in your hand, along with the file size, quantization and context guidance. Its estimates account for how iOS maps model files into memory. Keep storage in mind too: a model needs its file size on disk plus room for the download, and Noema's documentation suggests about 5 GB of headroom.
Frequently asked questions
How much memory does my iPhone have?
Apple does not list it. Reported figures: 4 GB for iPhone 11, 12 and 13; 6 GB for iPhone 12 Pro, 13 Pro, the 14 series and iPhone 15; 8 GB for iPhone 15 Pro, the 16 series, 17 and 17e; 12 GB for iPhone 17 Pro, Air and 18 Pro.
Can an iPhone run an 8B model?
On iPhones with 12 GB, such as iPhone 17 Pro, Air and 18 Pro, an 8B model at 4-bit is about a 5 GB file and is the size to plan for. On 8 GB iPhones it is too tight; use a 4B model. An independent user reported about 23 tokens per second from an 8B model on a 12 GB iPad with M5.
What is the best model size for an iPhone with 8 GB?
About 4 billion parameters at 4-bit, a file of 2.5 to 2.7 GB. That covers iPhone 15 Pro, the iPhone 16 series, iPhone 17 and 17e.
How big a model can a 16 GB Mac run?
A 9B model at 4-bit, around a 5.7 GB file, runs comfortably with room for macOS and a useful context. A 14B at 4-bit (about 9 GB) fits but leaves little headroom.
Why does a model fit on my Mac but not on an iPad with the same memory?
macOS lets an app use most of the machine's memory, while iPadOS and iOS limit how much one app may take. On a tablet or phone, plan for a model well under the installed memory.
Can I run a model larger than my device's memory?
Sometimes. Noema's Overfit runs compatible mixture-of-experts models by paging their experts from storage, at reduced speed. The other route is to run the model on a Mac and use it from your phone, which Noema does through Constellation.
Does quantization make a model worse?
Slightly. A 4-bit file is about a quarter of the full-precision size with a small loss in quality, which is why it is the usual default. Below 4 bits the loss grows; above it, files get larger for smaller gains.
About this guide
NoemaAI makes Noema, one of the apps described here, so read this as a maker's guide and not an independent review. What it says about Noema matches the documentation. What it says about other products comes from their own listings, sites, repositories and documentation, linked below, as they read on . These products change often. If something here is wrong or out of date, write to [email protected] and we will correct it.
Sources
- Wikipedia: list of iPhone models
- Wikipedia: list of iPad models
- Apple: compare iPad models
- Apple: iPad Pro specifications
- Apple: iPad Air specifications
- Apple: Apple Vision Pro specifications
- Apple: MacBook Neo specifications
- Apple: MacBook Air specifications
- Apple: MacBook Pro specifications
- Apple: iMac specifications
- Apple: Mac mini specifications
- Apple: Mac Studio specifications
- Apple: Apple Intelligence device list
- Apple Machine Learning Research: exploring LLMs with MLX and M5
- Qwen3.5-4B GGUF files (Unsloth)
- Qwen3.5-9B GGUF files (Unsloth)
- Qwen3-14B GGUF files
- Qwen3.8-27B GGUF files (lmstudio-community)
- LocalLLaMA: M5 iPad runs an 8B Q4 model