Offline AI: When Your Tools Need to Work Without Wi-Fi

Liquid AI's new small model is reportedly beating considerably larger rivals on real phone benchmarks. That sounds like a technical detail until you think about where a lot of Canadian work actually happens: a basement with no signal, a job site an hour outside town, a service van, a clinic in a community where the connection drops for an afternoon at a time. Cloud AI is useless in all of those, and it has been quietly excluding a lot of people from tools everyone else takes for granted.

Small models got good enough

A year ago the honest advice was that on-device models were a curiosity. They ran, they were private, and they were clearly worse in a way you noticed within about five minutes.

That is no longer the whole story. Models designed specifically for phones and laptops have improved sharply, and the current crop handle everyday tasks well enough that most people would not immediately spot the difference on a summary or a draft. They are still small models with small-model limits. The change is that the limits now sit in places you can plan around rather than everywhere at once.

Three reasons that actually justify it

Reason	Who this is
Connectivity is unreliable	Trades, field service, rural and remote work
Data should not leave the device	Health, legal, anything with personal information
The work cannot stop	Anyone who felt the last provider outage

That first row covers more Canadian businesses than the tech coverage suggests. A lot of the country works in places where the signal is genuinely bad, and every AI product demo assumes a good connection and a desk.

The limits, stated plainly

Small models are weaker at complex reasoning, long documents, and anything needing broad world knowledge. They cannot look things up, so nothing that depends on current information is possible. Heavy use drains a battery in a way people notice on a long day. And they need decent hardware, so the phone your business bought everyone in 2021 may not be the test you want to judge this on.

The useful mental model is bounded versus open-ended. Summarise these notes, translate this conversation, pull the fields out of this form, answer questions about this manual: all fine. Research this supplier, analyse this contract, tell me what changed in the regulations: not fine, and it will still produce an answer, which is the part to watch.

Offline is not automatically private

The privacy argument is strong and it deserves one check rather than an assumption. If the model runs locally and the app truly sends nothing home, your data stays on the device, which removes a whole category of question under PIPEDA and provincial rules.

But plenty of apps run a small model on the device while still syncing logs, analytics, or fallback queries to a server when a connection appears. Ask the vendor exactly what leaves the device and when, and get the answer in writing. The general approach in keeping AI use PIPEDA-compliant still applies, it just gets much easier to satisfy when the answer is genuinely nothing.

The outage case

There is a third reason that gets forgotten until the day it matters. A hosted AI service going down takes your workflow with it, and we went through that exercise in working out what stops when your provider does. An on-device model for the one or two tasks that genuinely cannot wait is a fallback that does not depend on anyone else being online.

That does not mean running everything locally. It means knowing which single task would hurt most during an outage, and having that one covered.

How to try it without a project

Pick one bounded task your team does away from a desk. Transcribing a site visit. Drafting a short report from voice notes. Answering questions about a manual loaded onto the device. Run it for two weeks on real work, on the hardware your staff actually carry.

Compare output against your hosted tool on the same inputs, and watch battery and speed. If you already run models on your own infrastructure for cost or control reasons, the reasoning is a close cousin of running local models in a Canadian business, just moved from a server to a pocket. And if you decide the cloud version is better for your work, that is a perfectly good outcome from a two-week test.

Frequently Asked Questions

What is offline AI?

It is an AI model that runs on the device in front of you rather than in a data centre, so it works with no internet connection at all. The model lives on the phone, tablet, or laptop, and nothing you type leaves the hardware. Until recently the small models that fit on a device were noticeably worse than the hosted ones. That gap has narrowed considerably, with new on-device models posting results that beat much larger rivals on phone benchmarks, which is what makes this worth a look now rather than next year.

Who actually needs it?

Three groups, and the first is bigger in Canada than people assume. Anyone working where connectivity is unreliable: rural service areas, basements, job sites, mines, remote clinics, long stretches of highway. Anyone handling information that should not leave the device, where local processing removes an entire category of privacy question. And anyone who needs a tool to keep working during an outage, since a hosted service going down takes your workflow with it. If none of those describe you, cloud AI is simpler and better.

What are the real limits?

On-device models are smaller, so they are weaker at complex reasoning, long documents, and tasks needing broad world knowledge. They cannot look anything up, which rules out anything requiring current information. They drain battery meaningfully during heavy use. And they need reasonable hardware, so an eight-year-old laptop will struggle. The realistic framing is that offline AI is very good at bounded, well-defined tasks and noticeably worse than a hosted model at anything open-ended or research-shaped.

Does offline mean private?

It removes one large risk and does not remove all of them. If the model runs locally and the app genuinely sends nothing home, your data stays on the device, which is a meaningful privacy improvement and simplifies a lot of PIPEDA and provincial obligations. But verify rather than assume: plenty of apps run a small model locally while still syncing logs, analytics, or fallback queries to a server. Ask the vendor what leaves the device, get it in writing, and check whether an offline mode is truly offline or just partially so.

How should we test it?

Pick one bounded task your team does in the field and try it for two weeks on real work. Good candidates are transcribing a site visit, drafting a short report from notes, translating a conversation, or answering questions about a manual you loaded onto the device. Compare the output against what your hosted tool produces on the same inputs, and pay attention to battery and speed on the actual hardware your staff carry rather than on a new test device.