Apple's small on-device language model is now in the Holidaymaker guest app. Here is how it works, where it falls short, and what changed in version 2026.9.1.
Most of the AI people use today runs a long way from them. A question is typed, it travels to a data centre, a very large model running on racks of servers produces a reply, and the reply comes back. For a development team building a guest app, that architecture brings three things nobody wants to hand a holiday park. A bill for every request. A dependency on signal. And somewhere in the middle, a server holding guests’ text that someone has to secure and explain.
On-device AI removes all three by moving the model onto the phone. Apple Intelligence includes a language model that lives on recent iPhones, iPads and Macs and runs on the device’s own chip. Apple’s technical report puts it at roughly three billion parameters. Parameters are, loosely, the numbers a model learns during training, and the big cloud models have many times more. The small size is a deliberate engineering trade. The model has to share storage with a guest’s photos and run without draining the battery by lunchtime.
The change that mattered came with iOS 26. Apple opened that model to app developers through the Foundation Models framework, which in Apple’s words offers AI inference that is free of cost, works offline and protects privacy. In practice, an app can open a session with the model, give it instructions and some text, and get a response back without leaving the app, with no account, no API key and no network call. Everything in the Holidaymaker app’s AI features is built on that foundation.
Privacy comes first, because it falls straight out of the architecture. Text goes into the model on the phone and the response comes out on the phone. No server in the loop logs what guests read, which leaves nothing to protect on their behalf and nothing awkward for a reception team to explain.
Cost is next, and it’s the reason the approach works for a park of any size. Cloud AI is priced per request. On a bank holiday weekend, with hundreds of families opening the app, those requests add up, and someone has to forecast them. On-device inference has no per-request cost to the park or the guest. Busy weekend or quiet Tuesday, the AI costs the park nothing.
Then there’s connectivity, which anyone who has worked on a park takes seriously. Wi-Fi at the far end of a touring field in August is a hope rather than a guarantee. Once Apple’s model has downloaded, it needs no connection at all. A guest outside their lodge with one bar of signal gets exactly the same result as one standing next to the router.
Finally, speed. With no round trip, a summary starts appearing almost as soon as the page opens.
A fair account has to include the limits. They’re real, and most of the decisions later in this piece exist because of them.
Device support. Apple Intelligence needs an iPhone 15 Pro or later, an iPad with an M1 chip or newer, or the iPad mini with the A17 Pro chip. Plenty of guests arrive with a perfectly good older iPhone, and Android can’t run Apple’s model at all.
The download. Even on a supported phone, the model has to download first. Apple asks for up to 8 GB of free storage on most supported devices and up to 14 GB on its newest models. A phone full of holiday photos may never have it.
Languages. Apple currently supports 16 languages, including English, French, German, Spanish, Dutch and Japanese. That covers a lot of guests, but not all of them.
Capacity. A small model works on a limited amount of text at a time, which developers call its context window. It also knows nothing about a park beyond what it’s handed. It’s very good at reading a page and saying what’s in it. It’s poor at anything needing outside knowledge or long chains of reasoning. Asked to summarise a welcome pack, it does a tidy job. Asked for the best pub in the village, it has nothing to go on.
Guardrails. Apple wraps the model in safety checks, and they can be overcautious. Apple’s own release notes acknowledged that some prompts wrongly return a guardrail error. Ordinary promotional copy can occasionally trip one. When that happens the model returns an error instead of text, so the app has to handle that cleanly.
Platform. Underneath everything, this is Apple’s technology on Apple’s devices. It’s the one limit no amount of engineering removes.
Apple Intelligence isn’t new to the Holidaymaker app. Content summaries and recommendation reasons first appeared in earlier releases, and those versions did what they were built to do. They also gave the development team something more valuable than a launch: real park copy, real guests and real phones to learn from.
Some lessons were about the model. Promotional copy is full of invitations, and a small model filling gaps will occasionally turn an invitation into an offer. Guardrails sometimes refused perfectly ordinary text. Some approaches that looked sensible on paper made results worse in practice.
Other lessons were about people. Guests wanted to know why AI appeared on one phone and not another. Some wanted one feature without the rest. Parks wanted a say over which content the model could touch at all, because a safety notice or the small print on an offer should never be paraphrased.
Version 2026.9.1 is the team’s answer to those lessons. It adds the per-feature settings and status line for guests, the content and record controls for parks, grounding and the promise check in the code, clearer labelling and disclaimers, the weather summary, and Siri and Spotlight support on iOS 27. Each is described below. Updating the app also replaces every summary written under the older rules, because the app’s instructions to the model are part of each summary’s fingerprint, so nothing produced before 2026.9.1 lingers on screen.
The development team set a simple rule for 2026.9.1. Every AI feature does one modest job, sits beside the park’s own content, and can fail without the guest losing anything. Four features are built on that rule, all on iPhone and iPad.
The first is detail-page summaries, for places, events, offers, walking routes, homes for sale and the welcome pack. When a description runs past 400 characters, roughly a long paragraph, and the phone’s language is supported, the app asks the model for a short summary and shows it in a card on the page. The card says the text was summarised by Apple Intelligence and notes that it may not be complete or accurate. It also points to the full details, which always sit underneath, untouched.
Generating text costs battery and time, so each summary is stored on the phone once it’s written. The stored summary is tied to a fingerprint, called a hash, of three things: the park’s source text, the version of the app’s instructions to the model, and whether that content is currently allowed to use AI. If any of those changes, the old summary is thrown away and a new one is written. In plain terms, when a park edits a description in the CMS, guests get a summary of the new copy, never a stale one.
One decision here went against the team’s first instinct. Models choose each next word with a little randomness, and that can be switched off so the same input always produces the same output. For cached summaries, that sounded ideal. In testing, though, it made the model reliably pick its safety-refusal path on some perfectly ordinary copy, and the whole summary was lost. So the app leaves Apple’s default variation in place. A summary that reads very slightly differently on two phones is a much better outcome than no summary.
The second feature, new in 2026.9.1, is the weather. The app passes the day’s forecast figures to the model and asks for two or three friendly sentences about the day ahead, instead of a column of wind speeds and percentages. It’s explicitly told not to suggest activities. A model that has never seen a park has no business recommending its crazy golf.
The third explains recommendations. When the app suggests something based on a guest’s interests, the model can add a sentence or two on why it might suit them. From 2026.9.1 it works only from that item’s own description, up to about 1,200 characters of it. The wording is framed as a possibility, never a promise, and it carries a disclaimer. If the item has no description, the model isn’t called at all, and the app shows its standard reason instead.
The fourth, also new in 2026.9.1, puts a park’s content into Siri and Spotlight on iOS 27. Guests can ask Siri about events, places, offers and more, and park content appears when they search their phone. The AI’s role here is easy to overstate, so precision matters. The sentence Siri reads aloud is assembled by the app from the search results using fixed wording, so it’s predictable and checkable. The only text the model generates in Siri is a short spoken summary of the guest’s own booking.
The clearest lesson from earlier versions was that an AI feature doesn’t earn trust unless people can see what it’s doing and turn it off. So in 2026.9.1, every guest has a fuller Apple Intelligence page in the app’s settings.
At the top is a single switch, Apple Intelligence Enabled, which turns everything on or off at once. It’s on by default, because every feature is labelled and optional, and one tap turns it off.
Below that is a status line. It asks Apple directly whether the model can run on this phone right now, and reports the answer in plain words. Either the phone is ready, or the device isn’t eligible, or Apple Intelligence hasn’t been turned on in the phone’s own settings, or the model is still downloading. That one line answers the question earlier versions raised most often: why does a friend’s phone show a summary when this one doesn’t?
Once the model is ready, guests can choose feature by feature, with separate switches for Summaries, Recommendation Reasoning, Weather Summaries and Siri Shortcut Summaries. If the phone is set to a language Apple’s model doesn’t support, the Summaries switch is greyed out with a short note explaining why. That’s better than the feature quietly doing nothing.
For guests who want it gone altogether, there’s Hide Apple Intelligence Features. With Apple Intelligence switched off, this removes every AI feature and prompt from the app, which then looks exactly as it would on a phone without Apple Intelligence.
Park controls are new in 2026.9.1. They live in the CMS and work at two levels.
The broad level is a switch for each type of content: places, the welcome pack, events, walking routes, offers, homes for sale and FAQs. Turning one off keeps that whole category away from Apple Intelligence, and from Siri and Spotlight too, because one switch governs all three. A park happy for its events to be summarised, but wanting its homes-for-sale listings left exactly as the sales team wrote them, needs only a single change.
The fine-grained level works record by record. Any single item can be excluded, which suits the pages where every word matters: a safety notice, the pool rules, the small print on an offer.
Both levels are checked together every time. Content only reaches the model when its category is allowed and the record itself isn’t excluded. And excluding something from AI never hides it from guests. The record still appears in full in the app, just without anything generated on top. Because the AI setting is part of that summary fingerprint, flipping it clears any existing summary at the next sync, with no manual clean-up.
Settings decide where the AI appears. A second layer, written into the app itself in 2026.9.1, decides what it’s allowed to say.
The first rule is grounding. The model is only ever given the park’s own copy. That copy goes in clearly marked off from the app’s instructions, with an explicit line telling the model to ignore any instructions inside it. That matters more than it sounds. Without it, a sentence in park copy that happened to read like a command could change what the model does. Developers call this prompt injection, and it’s one of the first things any team working with language models has to guard against.
The second is the promise check, which runs on every summary and every recommendation reason before a guest sees it. The app looks for a short list of words that imply an entitlement: free, complimentary, included, inclusive, no charge, no cost, unlimited, discount, voucher, gift, bonus and on the house. If the AI text uses one that the original copy doesn’t, the text is discarded. The guest sees no AI text, and the park’s copy stays exactly as written. This rule came straight out of earlier versions, where “come and have breakfast with us” once became a free breakfast that was never on offer. A small model filling a gap with something plausible is exactly the failure a careful team designs for, and 2026.9.1 is where that design landed.
The third is honesty about reach. All of this runs only on iPhone and iPad. Every guest on any phone still sees the park’s full content. The AI adds a layer for those who can use it and removes nothing for those who can’t.
One note on analytics, in the interest of precision. From 2026.9.1, when a park has analytics on, the app counts how each AI feature performs: whether it succeeded or failed, why it failed, and roughly how long it took. It also records which AI switches a guest has on. It never records the text itself. Nothing the guest reads and nothing the model writes leaves the phone. Those counts give the team a measured view of how the features behave across parks, to inform the next round of improvements.
Put the features, settings and code-level rules together and the benefits are quite concrete.
For guests, it’s help they can see and choice they control. A family planning the day gets the short version of a long page, a readable forecast and a reason behind each suggestion, with or without signal. Anyone who’d rather not can turn off one feature, all of them, or hide the lot, and AI text is always labelled.
For park teams, it’s fewer puzzled conversations. The status line explains why a phone isn’t showing AI before anyone reaches the desk. And when a guest asks about privacy, the answer is short and true.
For a park’s brand, it’s a voice that stays protected. The content and record switches decide what the AI may touch, so terms, safety information and sales copy can stay word for word. Grounding and the promise check mean anything the AI does say only repeats what the park has already said, and never offers something it hasn’t. If a summary ever reads oddly, the fix is in the source text, and the app rewrites the summary as soon as that text changes.
For a park’s budget, there’s nothing to add. The model runs on hardware guests already own, however busy the weekend gets.
And for every guest, on every phone, nothing is lost. The full copy is always there.
If it asks one thing of parks in return, it’s what good guest communication has always asked. Clear, accurate copy gives the model better material, and it gives guests a better page whether they ever see a summary or not.
On-device AI today is modest, and that’s arguably its strength. It runs on some phones, in some languages, and when it isn’t sure, it would rather return nothing than guess. That suits a guest app well. The developer’s job is to build around those edges carefully, learn from how the features behave in the real world, and give parks and guests clear control over how much of it they use. Version 2026.9.1 is the latest step in that process, not the last.
The HolidaymakerXP team will keep watching what Apple ships and adjust as the model improves, without loosening the rules above. Parks that have seen guests using these features, or have content they’d like kept well away from them, are warmly invited to get in touch.
If you would like to talk through how these AI features could work for your park, get in touch. Email hello@holidaymakerapp.co.uk or call 01305 542075 and we will happily have a chat.
See what the September 2026 app update brings for your guests: personalised recommendations, on-device AI summaries and one-tap WiFi Connect, plus a few things to set up first.
Read more
The Holidaymaker platform keeps your venue's content in one CMS and shows it in your branded app, on screens and in chat, so you update once and guests see it everywhere.
Read more
A branded guest app puts your venue in every guest's pocket, with search, one-tap Wi-Fi, offers, personalisation and informational notifications, all updated once in your CMS.
Read more