Turn On Airplane Mode and Your App's AI Feature Dies. It Doesn't Have To.

Nikola Nikolov — Thu, 01 Oct 2026 12:24:55 +0000

Why is that?

That's because the AI was never in the app. It's an HTTP call to someone's GPU, with your user's text in the request body and a bill at the end of the month.

For a lot of features it doesn't need to be. A 330 MB model runs on an iPhone 11. Here's how to put one inside your app.

NobodyWho, the open-source library i'm using below.

The whole setup

React Native / Expo:

npx expo install react-native-nobodywho

import { Chat } from "react-native-nobodywho";

const chat = await Chat.fromPath({
  modelPath: "hf:NobodyWho/Qwen_Qwen3-0.6B-GGUF:Q4_K_M",
});

const msg = await chat.ask("What is the capital of Denmark?").completed();
console.log(msg); // The capital of Denmark is Copenhagen.

That's it. The model downloads from Hugging Face on first run and is cached after that. From then on it works offline, with no API key, no server and no per-token bill.

Flutter, Swift and Kotlin look almost the same (Chat.fromPath, then ask). There's also Python for desktop and Godot for games.

What you can actually build with it

Summaries, tagging or rewriting text that shouldn't leave the phone (notes, journals, messages)
Features that have to work offline, on a plane or with bad signal
Structured output and tool calling. NobodyWho builds the grammar from your function signatures, so you don't write JSON schemas by hand
Voice: speech-to-text with Whisper and text-to-speech run locally too

What you give up

Being honest here, because this is where people get burned:

Model size. The Qwen3 0.6B model above is ~330 MB and runs on any phone. Bigger models need a lot more RAM, roughly 2× the file size free.
Device floor. iPhone 11 / 4 GB RAM or newer. On Android, around Snapdragon 855 with 6 GB RAM.
It's not a frontier model. A small local model is great at narrow jobs and bad at being a know-it-all.
Gaps. No web export yet, and no Windows ARM64.

So it won't replace a frontier model. But for a lot of app features you were about to send to an API, it's enough, and it still works in airplane mode.

Try it

Starter apps for Expo, React Native, Flutter and Swift, docs at docs.nobodywho.ooo.

Jev: new paradigm, or just logits with a better marketing team?

Nikola Nikolov — Fri, 25 Sep 2026 11:33:48 +0000

Jev is being pitched as a brand new kind of model: give it a fixed set of choices, get calibrated probabilities back in one shot.

but isn't that just a forward pass and a softmax? you can already do that fully on-device:

NobodyWho: open-source, drop-in local LLMs for basically any stack (Python, Swift, Kotlin, Flutter, RN, Godot)
Cactus: on-device inference for phones, already routes by model confidence
llama.cpp: the OG, grab the logits yourself

I get that the secret sauce is supposed to be calibration, and sure, raw logits from a small model aren't calibrated and usually have a favorite letter. but is that a real moat?