Which iPhones run Sandbook

Two requirements: iOS 18.2 or later, and a chip from the A14 generation or newer. On iPhone that means the iPhone 12 family and everything since, including the iPhone SE (3rd generation). On iPad it means iPad Air (4th generation) and later, iPad (10th generation) and later, iPad mini (6th generation) and later, and iPad Pro with M1 or later.

The reason is the voice itself. Sandbook does not stream audio from a server; it generates every sentence on the device with Kokoro-82M, run through Apple MLX on the GPU. MLX needs the GPU family Apple introduced with the A14 chip. Earlier iPhones do not have it, so on those phones the model cannot run at all, and Sandbook would rather say so than pretend. Other neural readers draw the line in different places; the survey of what each app asks of an older iPhone has the details.

The App Store's own device rules cannot express “A14 or later”, so an older iPhone may still be allowed to install the app. If it is, you can import and read books, and the app explains that voices need an iPhone 12 or later. That is not a useful reader, which is why the next section exists.

Sandbook's voices do not work on the iPhone XS, XS Max, XR, 11, 11 Pro, 11 Pro Max or SE (2nd generation), even on iOS 18. What does work on every one of them is built into iOS already, and it is free:

• Speak Screen. Turn it on in Settings, Accessibility, Spoken Content. Then swipe down with two fingers from the top of any screen and the iPhone reads the whole page, with a small controller for speed and skipping.
• Speak Selection. In the same settings. Select any text and tap Speak in the menu that appears.
• Better voices. Under Spoken Content, Voices, you can download an Enhanced or Premium voice for your language, which sounds noticeably better than the default and works offline once downloaded.

Each setting is walked through, with the current menu names, in how to make an iPhone read text aloud.

What Speak Screen does not do is behave like a book reader: there is no library, no importer for EPUB or PDF, and no reading position kept per file. The full comparison, including how to set it up, is on the Speak Screen alternative page.

The model is small. Kokoro-82M has 82 million parameters, and Sandbook's whole download is about 276 MB with the model and all 41 voices included, which is minuscule next to the speech models that live in data centres, and small enough to sit in a phone's memory and be used rather than merely stored.

MLX uses unified memory, so the weights are not shuttled back and forth between CPU and GPU on every step. On a phone, where memory bandwidth is the scarce resource, not copying things is most of the optimisation.

And real-time speech is a latency problem, not a throughput problem. Sandbook synthesizes sentence by sentence and starts playing as soon as the first sentence is ready, so the GPU only has to stay slightly ahead of a person listening at conversational speed, a target an iPhone 12 meets comfortably.

There is no lite mode, no reduced voice set and no lower quality tier. An iPhone 12 gets the full Kokoro-82M model at 24 kHz, all 41 voices (28 US and UK English voices plus Spanish, French, Italian, Brazilian Portuguese and Hindi), word-by-word highlighting, EPUB, PDF, web pages and scanned pages, the reading themes and fonts, and the full 0.5× to 2.5× speed range.

It also gets the same independence from the network. Because the model is on the phone, there is no buffering, no streaming, no failed request in a tunnel, and no data used while you read a book offline.

One limit applies to every device, new or old: real-time reading pauses when the screen locks. To keep listening with the screen locked, prepare a chapter's audio ahead of time first.