Which AI model is best for my iPhone?

Updated October 17, 2026 · 11 min read

It all comes down to your iPhone’s RAM. With 12 GB (iPhone 17 Pro, Air, 18 Pro), pick Gemma 4 E4B; with 8 GB (iPhone 15 Pro, 16, 17), Gemma 4 E2B or a compressed E4B; with 6 GB, a compressed Gemma 4 E2B; below that, Gemma 3 1B. Here’s how parameters, quantization, RAM and storage fit together.

The short answer: your iPhone, its RAM, the right model

What decides which model you can run is not storage or the chip, but memory (RAM). Apple does not show it in Settings or list it on its spec pages, so the figures below come from the benchmarks and teardowns compiled by Wikipedia. Find your model under Settings → General → About, then read the matching row.

iPhones older than the XS and XR can’t install iOS 17, the minimum for recent local AI apps such as DontKeep. On launch, DontKeep reads the device’s RAM, hides models that wouldn’t fit and highlights the one that does, so you don’t need to memorize this table.

The models at a glance

Every model below is a Gemma, Google’s family of open models, in the GGUF format read by the open source llama.cpp engine, which runs on the iPhone’s GPU through Metal. The score is Google’s published result on MMLU-Pro (university-level questions across about fifteen subjects). It isn’t an absolute grade, but it’s a fair way to compare models with each other.

Google doesn’t publish separate scores for the compressed versions: the figure is the original model’s, and the real result sits a little lower. The optional “images and dictation” module adds to the file size (see below). The model cards are also shown in the Models section of the home page.

1B, 4B, E2B, E4B: what the names mean

“B” stands for billion: it’s the number of parameters, the internal settings learned during training. Gemma 3 1B has about one billion, Gemma 3 4B about four. As a rule, more parameters means more nuance, more knowledge and better instruction-following, but also more memory.

Gemma 4 adds a letter: the “E” in E2B and E4B stands for effective parameters. According to Google’s model card, E2B has 2.3 billion effective parameters (5.1 billion including embeddings) and E4B has 4.5 billion effective (8 billion in total). The gap comes from a technique called Per-Layer Embeddings: each layer gets its own small vocabulary table, which is large but only looked up, not computed. The model “thinks” like a 2B or 4B model while taking more space than you’d expect, which is why an E4B, even compressed, needs a recent iPhone.

Between the two, the difference shows on demanding tasks: E4B (69% on MMLU-Pro) reasons better, structures long texts better and makes fewer factual mistakes than E2B (60%). For a quick question or a short email, E2B is plenty and answers faster. Both understand text, images and audio, and were trained on more than 140 languages.

4-bit vs 2-bit: quantization and what you lose

Originally, each parameter is stored in 16 bits. Google says that at this precision E4B would need about 18 GB of memory for its weights alone, far beyond a phone. Quantization rounds those numbers so they fit in fewer bits. At 4-bit (Q4_0), E4B drops to about 4.5 GB and E2B to 2.9 GB, usually with a small loss in quality: it’s the standard for local AI.

Going down to 2-bit is trickier, since rounding that hard normally damages a model. Gemma 4’s “mobile” versions limit the damage with QAT (quantization-aware training): the model is trained knowing it will be compressed, and learns to compensate. The versions DontKeep uses come from these QAT checkpoints, converted to GGUF by Unsloth.

In practice, a 2-bit E4B keeps most of its reasoning ability, but it’s a little less accurate than the same E4B at 4-bit: occasionally clumsier wording, slightly more slips on details, numbers or less common languages.

DontKeep makes this choice for you: on a 12 GB iPhone it offers the full 4-bit version; on 8 GB, the compressed E4B (the full one wouldn’t fit); on 6 GB, the compressed E2B. Only one version of each model is shown, the one that suits your device.

RAM vs storage: two different things

Storage (128 GB, 256 GB…) keeps the model file on the iPhone. RAM runs it: while you chat, the whole model has to be loaded into memory. A 512 GB iPhone with 6 GB of RAM still can’t run a model that needs 8.

For storage, plan for the file plus the optional module: about 5.2 GB for 4-bit E4B with its module, 4.2 GB for compressed E4B, 3.4 GB for E2B, 3.2 GB for compressed E2B, under 1 GB for Gemma 3 1B. Leave some free space for iOS, too. You download once; after that, everything works offline.

Context: the conversation’s memory

The context is how much text the model keeps in mind at once: your messages, its replies, a pasted document. It’s measured in tokens, chunks of words (in English, roughly 1,000 tokens for 750 words). Gemma 4 accepts up to 128,000 tokens in theory, but every token kept takes RAM in a cache (the KV cache), on top of the model itself.

On a phone, that’s a trade-off. DontKeep sets the default context to 16,000 tokens on 12 GB iPhones, 8,000 on 8 GB and 4,000 below, and lets you change it in settings. Beyond that, the model “forgets” the start of the conversation. For summarizing long documents, a 12 GB iPhone is a real advantage.

Why a model that’s too big makes the app quit

iOS doesn’t give one app all the RAM: the system, the display and other apps keep their share. Even with the special permission Apple grants to memory-hungry apps (Increased Memory Limit, which DontKeep uses), an app that goes over its limit is terminated by the system on the spot, with no error message: it looks like a crash. The model, its context cache and, if loaded, the image module all have to fit under that ceiling. That’s why a 4.6 GB file needs a 12 GB iPhone, and why it’s best to let the app filter what fits your device.

How fast will it be?

To write each word, the model has to read a large part of its weights from memory, so speed mostly depends on model size and the chip’s memory bandwidth. On the same chip, a model half the size answers roughly twice as fast. Expect a short loading time when you open the app, then text that streams in: on a recent iPhone with the recommended model, reading generally keeps pace with writing without a frustrating wait; Gemma 3 1B and 270M are much faster, with simpler answers. Long sessions warm the phone, which can slow generation a little, and use battery.

The images and dictation module (mmproj)

The base model only reads text. To show it a photo, a screenshot or a scanned document, you need an extra module called mmproj (multimodal projector), which turns the image into data the model understands. For Gemma 4 E2B and E4B, the same module also handles audio, enabling voice dictation processed on the iPhone. It weighs 0.56 to 0.99 GB depending on the model, is optional, and also uses RAM when active. Gemma 3 4B reads images but not audio; Gemma 3 1B and 270M are text-only.

“Uncensored” models: what are they?

Google’s models decline some requests. Independent developers publish modified versions with those refusals removed: abliterated (you find the “direction” tied to refusal inside the model and erase it, a method popularized by Maxime Labonne) or Heretic (a tool that automates the process). DontKeep has a section for such models, turned off by default.

• They are not official releases: the original publisher neither tests nor endorses them, and their quality can be slightly lower.
• No refusals doesn’t mean reliable: the model can state false or dangerous things with confidence. Always double-check, especially for health, legal or financial matters.
• You remain responsible for how you use them: the law and the original model’s terms still apply. These models are not suitable for minors.

Our recommendations by iPhone

12 GB iPhones (17 Pro, 17 Pro Max, Air, 18 Pro, 18 Pro Max)

Go with Gemma 4 E4B at 4-bit, plus the images and dictation module if you have the space. It’s the best model you can run on an iPhone today, with the largest context (16,000 tokens). Add Gemma 3 1B if you want a second, very fast model for small questions.

8 GB iPhones (15 Pro, the whole 16 range, 17 and 17e)

Start with Gemma 4 E2B at 4-bit: fast, dependable, the best balance. If you write a lot or ask reasoning questions, try the compressed Gemma 4 E4B: smarter, a bit slower and a bit less precise than at 4-bit.

6 GB iPhones (12 Pro, 13 Pro, the 14 range, 15 and 15 Plus)

Compressed Gemma 4 E2B is the default: the latest generation in 2.2 GB. Gemma 3 4B QAT is a solid alternative, especially for reading images. Keep conversations short (4,000-token context) and close heavy apps if the app quits.

4 GB or less (XS, 11, 12, 13, SE)

Gemma 3 1B on 4 GB, Gemma 3 270M on 3 GB. They handle simple questions, rewording and short translations, but won’t replace a large online model for complex tasks. Be realistic: they often get facts wrong.

Which model for which use?

• Writing (emails, letters, delicate messages): E4B if your iPhone allows it, otherwise E2B. Models under 2 billion parameters write correctly, with less nuance.
• Summarizing documents: context size matters as much as the model. On 12 GB, E4B with 16,000 tokens can summarize a few dozen pages; on 8 GB, split long texts into parts.
• Quick questions, rewording, translating a sentence: E2B, or Gemma 3 1B for speed.
• Reading a photo, a receipt, a screenshot: a Gemma 4 or Gemma 3 4B with its image module.
• News, prices, recent events: no local model, since their knowledge stops at their training date (January 2025 for Gemma 4) and they have no internet access.

Frequently asked questions

How do I find out how much RAM my iPhone has?

iOS doesn’t show it. Note your model name (Settings → General → About) and check the first table. A local AI app such as DontKeep reads the device’s memory directly and suggests the right model.

Can I install several models?

Yes, as many as your storage allows, but only one runs at a time: there isn’t enough RAM for two loaded models.

Can an 8 GB iPhone run the full E4B?

Not reliably: with the context cache, the system would close the app. That’s exactly what the 2-bit compressed version is for.

4-bit E2B or 2-bit E4B on an 8 GB iPhone?

E2B for speed and consistency, compressed E4B for tasks that need more thinking. The easiest way is to try both on your usual questions.

Is a local model as good as ChatGPT?

No, online models are far larger. A local model is still very useful for writing, summarizing or explaining, with one decisive advantage: your messages are processed on the phone, without being sent to a server.

Do I need an internet connection?

Only to download the model, once. After that, everything works in airplane mode.