On-device AI · 6 min read

What is a local AI e-reader, and why would you want one?

Most reading apps that "explain" words for you do it somewhere else. A local AI e-reader does it on the device in your hand. That single difference changes what the app can do on a plane, what it costs to run, and what anyone can ever learn about what you read.

The short definition

A local AI e-reader is a reading app that runs a language model directly on your phone or tablet, instead of sending text to a server. When you tap a word, the explanation is generated on the device, from a model file stored on the device, using the device's own processor.

Nothing about the sentence you tapped leaves the hardware. There is no API call, no request log, no account attached to the lookup.

Why the distinction matters

1. It works when the network does not

Cloud lookups fail in exactly the situations where reading is most appealing: on a plane, on the underground, in a hotel with hostile Wi-Fi, abroad without a data plan. A local model has no opinion about your connection. The dictionary in your pocket does not need a signal, and neither should the thing that replaced it.

2. Your reading history stops being data

The words a person looks up are unusually revealing. They expose which language you are learning, roughly how good you are at it, what you read late at night, which subjects you are quietly researching. In a cloud product, that stream is a server-side record by default — not because anyone is being sinister, but because request logs are simply how servers work.

When the model runs locally, that record does not exist to be kept, leaked, subpoenaed or sold. This is a stronger guarantee than a privacy policy, because it is structural rather than promissory.

A promise not to look at your data depends on the company keeping it. An architecture that never receives your data does not.

3. No metering, no per-lookup cost

Cloud inference costs the developer money on every single tap, which is why cloud-backed reading apps tend to arrive with subscriptions, credits, or daily limits. Local inference costs the developer nothing at run time. You can tap two hundred words in an evening and the economics do not change.

What you give up

Local AI is not free of trade-offs, and anyone claiming otherwise is selling something.

• Storage. A useful model is measured in gigabytes, not megabytes. That is a real cost on a phone that is already full.
• Raw capability. A model small enough to run on a phone is not the largest model in the world. For defining a word in context it is more than sufficient; for writing you an essay it is not the tool.
• Hardware floor. Running a language model locally is demanding. Older or low-memory devices either cannot do it or do it too slowly to feel good.
• Heat and battery. Sustained inference warms a phone. Well-built local apps pace their work rather than running the processor flat out.

How it works, briefly

The model ships as a quantised file — usually in GGUF format — which is a compressed representation of the model's weights. Quantisation trades a little accuracy for a large reduction in size and memory, which is what makes a phone-sized model possible at all.

A runtime such as llama.cpp loads that file, memory-maps it, and runs inference against the device's CPU or neural accelerator. The app hands it the tapped word plus the surrounding sentence, and streams back the answer token by token — which is why a good local reader shows the explanation appearing progressively rather than all at once.

What to look for

If you are evaluating a reader that claims to be local, these questions separate the real thing from marketing:

• Does word lookup still work in aeroplane mode? This is the honest test.
• Is there an account? Local processing and mandatory sign-up sit oddly together.
• What does the privacy policy say specifically leaves the device — and does it name the model?
• Are there per-lookup limits or credits? Those usually imply a metered backend.
• Does it work with your own files, or only with a store you have to buy from?

Where ClickBook sits

ClickBook is built this way on purpose. The model runs on the device, your books are files you already own, there is no account, and the text you read is not sent anywhere for cloud processing. Reading and word lookups work with the network off entirely.