Learn · Offline AI for field apps

Can an AI assistant in a field app work offline, with no signal?

Yes, for tasks a small model can be taught, if the model and the data it needs are already on the phone. A test build (31 Oct) of our first build answered a stock question in 4.25 s on a Galaxy A05s with 4 GB, airplane mode on. The first turn of that run took 13.4 s (right after a print preview); still open. That was pharmacy software in a device lab. No field worker has used it. FTL

Measured on a real phone in Firebase Test Lab (FTL). Test phone: Galaxy A05s, 4 GB, Snapdragon 680. Android 15, airplane mode, run PASSED. Test build, 31 Oct; no later merged build has passed a whole-app run yet. Every row with its conditions is on our evidence page. First build: Nesayo (pharmacy software, nesayo.com). The live Nesayo product uses cloud AI; the offline edition runs our model inside the phone, is tested on real phones in a device lab (Firebase Test Lab), and a shop pilot is next. We have no build for a collection or health-worker app. Published 6 Nov 2026.

Before the route

What has to be on the phone before the agent loses signal?

The model, room for it in memory, and the data it will be asked about. No signal means none of these can be fetched at the door.

Model size FTL
489 MB. qwen3-0.6b-v18-Q6_K.gguf. This file has to reach the phone while it still has a connection
Model memory FTL
673 MB. The model in memory during the whole-app run, on a phone with 4 GB in total
First launch FTL
18.8 s. cold_start_ms 18786. The wait when the app starts cold
Bill from a sentence FTL
5.76 s. Same whole-app run. The closest task we have measured to writing a record from one line

But the model is the smaller half of the job. It can only work with what the app has stored on the phone. In the offline build, bills, tax and stock are computed by code from local data, not by the model. Therefore an offline assistant is only as current as the app's last sync. The speed-bench figures are a separate run, on our benchmarks page.

Illustration · not built, not measured

What would a build for a loan collection app do?

Illustration only. This is what a build for your app would aim to do; none of it exists today, and none of it has been measured.

  1. Before the route, the app stores the day's visit list and each borrower's dues on the phone.
  2. In a village with no signal, the agent asks what a borrower owes. The model picks the app's lookup, code reads the stored record and does the sums, and the model words the reply.
  3. The agent says the borrower will pay on Friday. The model drafts the visit note, the agent approves it, and the app holds it until signal returns.
  4. A payment made at the branch that morning is not on the phone. The assistant cannot know it until the next sync, so the screen has to show when the data was last updated.

To find out whether this works, we would use our method on your app: map its tools, build a sandbox, freeze an exam from real traffic, teach a small open model from a larger teacher with checks, add code guards, quantise, and pass real-device gates on your users' own phones.

Offline build · SIMULATED

What fell short offline in our first build?

Reading photos of documents, the row a field app should read first. These rows are SIMULATED or code, not measured on a phone; they are not device proof.

Features working offline SIMULATED
95 of 106. 95 PASS, 4 DEGRADED, 5 FAIL (photo/PDF stock entry, prescription scan), 2 NEEDS-INTERNET. Older test build, not the current one. Many PASS rows only open a page or record a status
Spoken questions on the device SIMULATED
whisper.cpp voice input in the offline build. We make no accuracy claim for voice on the phone
Languages code
English, Hinglish, Hindi. Any other language your agents use would need its own check

If your agents photograph documents, this does not do that today: a vision model is on our roadmap as research, not available.

What does this result not establish for field work?

  • Patchy signal. Airplane mode is a fixed no-network state; this result does not establish how an app behaves when signal comes and goes.
  • A day on the road. Energy use was not measured, so nothing here speaks to battery life on a route.
  • Other phones. These are results from one phone model, the Galaxy A05s.
  • Your tasks. Accuracy needs an exam on your app's tasks. Our one exam result is on an earlier version: 0.6B vs its 8B teacher, +13 pts (95% CI +4.0 to +24.4); v7, not the shipped v18; frozen real test, 40 shops, shop-clustered, scored on a server. 16 of the 40 exam shops are QA-script traffic. EVAL
  • Field use. No paying customer, and no pilot running. A shop pilot is next.

The limits on our other results are on the same evidence page.

Questions

How do field workers on cheap Android phones use AI when the internet is slow?

With the model on the phone, the answer does not wait for the network. The only whole-app run that has passed, a test build of 31 Oct: a stock answer in 4.25 s on a Galaxy A05s with 4 GB, airplane mode on; the first turn of that run took 13.4 s (right after a print preview); still open. That is one lab phone, not a field worker on a route. FTL

Is there offline AI for ASHA and health-worker apps in rural India?

We do not have one today. Our first build is pharmacy software, and no health worker has used it. A build for a health-worker app would need an exam on that app's own tasks and its own measurements on the phones those workers carry. This page does not survey other vendors.

Does any data leave the phone?

In our first build, the AI and the shop's data stay on the phone, with one exception in the next test build: Check my plan, on the owner's tap, sends sign-in details to nesayo.com to read plan status; no shop data, no AI. No egress check has run on a phone yet. Whatever your app syncs when signal returns is your app's own data flow; your compliance team decides the rest.

Does this remove the cloud AI bill?

There is no per-question cloud bill, because no cloud AI answers the question. We have not measured a saving: the cloud cost baseline is not measured.

Is your field app a fit?

Probably, if your agents repeat a few tasks on phones you can name, and the data those tasks need can sit on the phone. Probably not, if you need a general chatbot, answers that depend on live data, reading photos of documents, or iPhone: iPhone is on the roadmap, and today we run on Android and a tested desktop build.

Tell us what your app does, one task you want to work offline, your agents' phones, and where the signal drops. Our design-partner page says what working with us involves.

Bring your software