Next-Word Prediction

Next-word prediction suggests what you are likely to type next — before you press a single letter. It draws on two sources: a built-in phrase model that ships with the keyboard (identical on every device, nothing personal in it), and — only if you turn on on-device learning — phrases learned from your own typing. It is fully on-device and on by default (since v2.0); the built-in part works even with learning off.

Quick Summary

What	Description
Purpose	Suggest the next word before you type a letter
Access	Settings > Input Behavior > Word Prediction > Next-Word Prediction
Default	On (since v2.0; an install where you had switched it off keeps it off)
Data source	Built-in phrase model (English: a model built from public text corpora; French, German, Italian, Portuguese, Spanish: a smaller built-in list of common word pairs). With Learn From My Typing on, your own learned phrases come first. Nothing downloaded, nothing leaves the device, and next-word itself records nothing

How It Works

Next-word suggestions fill the suggestion bar at moments when it would otherwise be empty. They come in two tiers:

• Built-in (always available) — for English, a phrase model that ships inside the app: about 75,000 common word pairs over 15,000 words, built from the Leipzig Corpora Collection (English web text, 2018) and Tatoeba example sentences. For French, German, Italian, Portuguese and Spanish a smaller hand-picked list of common word pairs plays the same role. This tier contains nothing about you, so it works with on-device learning off, with Context-Aware Predictions off, and in private/incognito fields.
• Learned (optional) — with Learn From My Typing on (Privacy & Data) and Context-Aware Predictions on, CleverKeys also privately learns which words you use after which (word pairs and three-word sequences). Those learned phrases always come first; the built-in tier tops up any slots they leave free.

Next-word prediction itself never records anything — it only reads. Learning happens (or not) according to your Privacy & Data settings, exactly as for ordinary typing.

Suggestions appear:

• After you finish a word with a space — the bar offers up to 3 likely next words.
• After you tap a suggestion — the bar refills from the new context, so you can chain whole familiar phrases by tapping suggestions without touching the letter keys.
• After a swipe — your swipe’s alternate readings stay in the bar (so you can still correct the swipe), and up to 2 next words are appended after them.
• When you tap into text — parking the cursor after a word offers continuations of the text just before the cursor (including text typed in an earlier session), similar to other modern keyboards. The text is read only for that lookup and is not stored.

Learned suggestions only appear when the keyboard is reasonably confident: a phrase must have been seen at least twice, and the continuation must be likely (at least a 5% chance given your history). Built-in suggestions are only ever used to fill slots your own data did not — they can never push a learned suggestion aside. An empty bar is still normal — showing nothing beats showing noise. Every suggestion, built-in or learned, is filtered against your dictionary, so a typo you made twice will not haunt the bar.

Walkthrough: typing a sentence

Say you often type “I want to go home”, and next-word prediction is on. With learning off, only built-in continuations appear — for example after want: to · you · a. The learned entries below additionally need Learn From My Typing on and the phrase typed at least twice before.

Tap typing:

1. Type I and press space. The bar shows learned continuations, for example want · am · think.
2. Tap want. The word commits with a space, and the bar refills: to · a · more.
3. Tap to, tap go… you can ride your own common phrases tap by tap.
4. Start typing a letter instead — the next-word suggestions vanish and normal letter-by-letter predictions take over.
5. Press backspace while no partial word exists — the next-word suggestions dismiss.
6. End a sentence with . ? or ! — the phrase context resets, so nothing carries across sentences.

Swipe typing:

1. Swipe “want”. The word inserts automatically, and the bar shows the swipe’s alternate readings (want · went · wart) — tapping one of those replaces the inserted word.
2. A moment later, up to 2 next words are appended after the alternates: want · went · wart · to · more. Tapping to appends “to” after “want” — it does not replace your swipe. Both behaviors live in the same bar; long-press any entry to see which kind it is.

Seeing why a word was suggested: long-press any next-word suggestion. A small sheet shows its source (“Next-word prediction”) and, for a word learned from your own typing, the statistics behind it, e.g. After “want to”: seen 14×, 63%. Built-in continuations say After “the”: common continuation (built-in, not learned) instead of quoting statistics they do not have. You can also enable colored origin dots for every suggestion under Settings > Advanced > Suggestion Origin Markers; with a screen reader (TalkBack) a marked suggestion is read as the word followed by its origin, e.g. “to, Next-word prediction”.

When suggestions will NOT appear

Condition	Why
Feature toggle switched off	Your choice (it is on by default)
Word prediction switched off	Next-word is part of word prediction
Password fields	Secure mode
Terminal (Termux)	Terminal input is not prose
An autocorrect-undo, add-to-dictionary or “Prefer … when swiping?” prompt is showing	Prompts take priority
Start of a sentence (after . ? !)	There is no previous word to continue
No learned phrase above the confidence floor, and no built-in continuation for the last word	Empty bar by design. The English model covers about 14,500 previous words; the French, German, Italian, Portuguese and Spanish lists only the most common opening words

Only the learned tier is switched off — the built-in tier keeps working — when:

Condition	Why
“Learn From My Typing” off (Privacy & Data; the default on new installs)	Your learned phrases are neither recorded nor read
Context-Aware Predictions off	That setting controls the learned phrase model
Private/incognito fields (apps that request no personalized learning, e.g. browser private tabs)	The app asked the keyboard not to personalize; built-in suggestions are not personal, just like ordinary word completions there

Configuration

Setting	Default	Description
Next-Word Prediction	On	The feature toggle (Input Behavior > Word Prediction)
Learn From My Typing	Off on new installs (upgrades keep it on)	Master privacy switch (Privacy & Data) — needed only for the learned tier
Context-Aware Predictions	On	Learns phrase patterns — needed only for the learned tier
Context Source	Both	Which phrase model boosts predictions: built-in, your learned patterns, or both
Personalization Strength	1.0	How strongly your word usage boosts predictions (0 = off, 2 = double)
Suggestion Origin Markers	Off	Colored dot per suggestion showing which engine produced it (Advanced)

Tips and Tricks

[!TIP] If you turn on Learn From My Typing, give it a few days. The built-in model covers common phrases from the start, but the confidence floor on your own data (a phrase must be seen at least twice) means the suggestions get noticeably more personal the more you type.

[!TIP] You can inspect and delete anything the keyboard has learned under Settings > Input Behavior > Learning & Data — browse learned phrases and words, delete individual entries, or forget everything.

Related Features

• Autocorrect & Predictions - The main prediction pipeline
• Privacy Settings - The master learning switch and data controls
• Input Behavior Settings - Where the toggles live

Technical Details

See the Next-Word Prediction Technical Specification.

Explicit apostrophe suffixes (development build)

The assignable Append ‘s and Append apostrophe commands attach exact text to the last verified word. With learning enabled, a successful suffix replaces that word’s own learning increments and context; immediate Backspace restores them through a fresh receipt. Moving the cursor, changing fields or clearing learned data expires ownership. These commands do not guess grammatical possession or alter contraction ranking. See per-key actions.