We make Pinku, one of the apps that run these models on the iPhone. The iPhone numbers below are not ours: they come from a public benchmark by an independent developer, linked in each table. Our own measurement was made on a Mac with an M3 chip, not on an iPhone, and is labelled that way. We have not measured Pinku on an iPhone yet; this page will be updated when we do.
The short answer
On an iPhone 17 Pro, small models of 0.6 to 1.7 billion parameters at 4 bits generate 37 to 62 tokens per second and show the first word in 0.16 to 0.7 seconds, with no internet. Llama 3.2 1B, Qwen3 0.6B and LFM2.5 1.2B are the fastest; Gemma 3 1B is the slowest to start, and gets slower as the conversation grows. In a chat app expect about 10 to 15% less than these figures.
Speeds on an iPhone 17 Pro (third-party benchmark)
| Model (4-bit, MLX) | Size | Tokens per second | Time to first token |
|---|---|---|---|
| Qwen3 0.6B | 351 MB | 62.2 | 163 ms |
| LFM2.5 1.2B Instruct | 663 MB | 59.7 | 244 ms |
| Llama 3.2 1B | 713 MB | 58.1 | 253 ms |
| Qwen3 1.7B | 984 MB | 39.5 | 360 ms |
| Gemma 3 1B (QAT) | 772 MB | 37.1 | 693 ms |
iPhone 17 Pro (A19 Pro, 12 GB, iOS 26.3 RC), median of 10 runs with a prompt of about 279 tokens and 256 tokens of output. From Ricky Takkar, How Fast Are On-Device LLMs on iPhone 17 Pro and iPad Pro?, 8 February 2026, measured in his app Russet with Apple's MLX framework. The author says the figures are the ceiling of the chip: in real use, turning tokens into text and updating the screen take roughly 10 to 15% off.
What else that test found
- Longer conversations are slower on iPhone. Qwen3 0.6B went from 69.7 to 43.4 tokens per second between the short and the long prompt. An iPad Pro M5 with the same memory was 1.2 to 2.2 times faster and barely slowed down.
- Gemma 3 1B is slow to start. Its time to first token grew from 407 ms with a 47-token prompt to 1,330 ms with a 690-token prompt.
- Bigger models did not fit. Qwen3 4B and Llama 3.2 3B were left out because they went past the memory iOS allowed that app on a 12 GB iPhone.
- 4 bits beat 6 bits. LFM2.5 at 6 bits was 21 to 27% slower than at 4 bits.
Our own test, on a Mac M3 (not an iPhone)
On 23 October 2026 we timed Pinku itself on a Mac with an M3 chip, with nothing else heavy running, to choose its default model. These are numbers from inside the app, with the screen updating, and they are from a Mac:
| Model | First text on screen | Tokens per second |
|---|---|---|
| Llama 3.2 1B | 0.13 to 0.16 s | 34 to 52 |
| Gemma 3 1B | 0.5 to 8 s, longer with each question | not recorded |
The pattern matches the iPhone benchmark: Gemma 3 1B takes longer to start, and the delay grows as the chat gets longer. That is why Pinku switched its default model from Gemma 3 1B to Llama 3.2 1B in version 1.2.
Apps that run these models offline on iPhone
| App | US ratings (30 Oct 2026) | What to know |
|---|---|---|
| Locally AI by LM Studio | 1,872 (4.7) | Runs models with MLX on iPhone, iPad and Mac |
| Enclave | 1,110 (4.7) | Local models; Pro adds cloud models |
| Private LLM | 741 (4.1) | Paid upfront ($4.99 in the US); works with Siri and Shortcuts |
| PocketPal AI | 165 (4.0) | Open source, runs any GGUF model with llama.cpp, has a built-in Benchmark page |
| Pinku | 12 (4.7) | Runs models with MLX on iPhone, iPad and Mac; its catalogue has all five models in the table above, at 4 bits; chats stay on the device |
Ordered by number of ratings in the US App Store. The speed of a model depends on the chip, the model and the runtime (MLX or llama.cpp) more than on the app, so the same model should run at a similar speed in apps that use the same runtime.
Pros and cons of running a model on the iPhone
- Pros: works in airplane mode; your text is processed on the phone; no wait for a server; small models start answering in under a second.
- Cons: small models know less and make more mistakes than ChatGPT or Gemini; no web search; long chats get slower; generating text warms the phone and uses battery.
Who it is for
- People who want quick answers, rewrites or translations with no connection: a 1B model is enough.
- People who do not want a conversation to leave the phone.
- Not for research or facts that must be right: use a cloud model with web search, and check it.
Step by step
- Check your storage. Small models are 0.4 to 1 GB. Settings, General, iPhone Storage shows how much room you have.
- Install an on-device chat app. Locally AI, Enclave, Private LLM, PocketPal AI or Pinku.
- Download a 1B model on Wi-Fi. Llama 3.2 1B, Qwen3 0.6B or LFM2.5 1.2B were the fastest on iPhone in the benchmark above.
- Turn on airplane mode and ask something. If it answers, the model is running on the iPhone.
- Measure your own phone if you want numbers. PocketPal AI has a Benchmark page that reports tokens per second and memory, and an optional AI Phone Leaderboard to compare devices.
- Start a new chat when it slows down. Speed drops as the conversation gets longer, so a fresh chat is faster.
Related
To see which models fit in your iPhone's memory, use which AI model fits my iPhone or Mac. For how on-device models work, read how local AI works on iPhone, and for a trip, how to use AI offline on a plane.
Sources
- Ricky Takkar, How Fast Are On-Device LLMs on iPhone 17 Pro and iPad Pro?, 8 February 2026, read 30 October 2026.
- Pinku, our own timing on a Mac M3, 23 October 2026 (app development log).
- PocketPal AI on GitHub: llama.cpp, GGUF and the Benchmark page.
- US App Store pages of each app, ratings read 30 October 2026.
- MLX, Apple's array framework for Apple silicon.
Questions people ask
How fast is Llama 3.2 1B on iPhone?
About 58 tokens per second, with the first token in about 0.25 seconds, on an iPhone 17 Pro at 4 bits with MLX (Ricky Takkar's benchmark, February 2026). Expect 10 to 15% less in a chat app.
Which local AI model is fastest on iPhone?
In that benchmark, Qwen3 0.6B (62 tokens per second), LFM2.5 1.2B (60) and Llama 3.2 1B (58). Gemma 3 1B and Qwen3 1.7B were slower, at 37 and 40.
Why is Gemma 3 1B slower than Llama on iPhone?
It takes longer to read the prompt before answering: about 0.4 seconds with a short prompt and 1.3 seconds with a long one on an iPhone 17 Pro, against about 0.25 seconds for Llama 3.2 1B.
Can an iPhone run a 3B or 4B model?
It depends on its memory and on the app. In the iPhone 17 Pro benchmark, Qwen3 4B and Llama 3.2 3B were left out because they went past the memory iOS allowed the app. Smaller models fit on most recent iPhones.
Does a local AI model on iPhone need internet?
Only to download the model once. After that it answers in airplane mode.
How do I measure the speed on my own iPhone?
PocketPal AI has a Benchmark page that reports tokens per second and memory use for the model you load.