Micro Models > SLMs > LLMs

let’s talk about running ai on phones.

everyone’s obsessed with making models bigger. more parameters, more data, more gpu. but for mobile agents, the ones that actually do stuff on your phone, bigger is the wrong direction entirely.

the numbers

	micro models	SLMs	LLMs
size	~80MB	~2GB	~40GB
cost	$0/action	$0.01/action	$0.03/action
runs on	on device	on device (barely)	cloud only

a micro model fits in your phone’s ram the way a photo does. an LLM needs a data center.

why send it to the cloud

every time your phone agent hits a cloud api, that’s latency. that’s a network dependency. that’s a privacy question. that’s a cost.

tap a button. 200ms round trip to a server. open an app. another 200ms. scroll down. another one. chain five actions together and you’re waiting a full second just for the model to think ² 2: and that’s assuming good connectivity. on spotty networks or in regions with high latency to cloud providers, it’s even worse. . on someone else’s computer.

micro models run locally. zero latency. zero cost. zero privacy concerns. the action happens before your finger leaves the screen.

tapping a button doesn’t need 70B params

here’s what mobile agents actually do most of the time. tap a button. type some text. scroll to an element. switch apps. read what’s on screen.

that’s it. pattern matching. spatial reasoning on a 6-inch screen. you don’t need a model that can write poetry and debate philosophy to find a button and tap it.

a 50M parameter model trained specifically on mobile ui interactions will outperform a 70B general model at these tasks ³ 3: we’ve seen this pattern before. specialized small models beat general large models in vision, speech, and translation. mobile ui is next. . every time. because it’s built for exactly this.

app-specific micro models

the future of mobile agents isn’t one massive model doing everything. it’s a swarm of tiny models, each trained for a specific app ¹ 1: i’m actively working on this. app-specific micro models trained to take decisions and perform actions for individual apps. paper coming soon. .

think about it. every app has its own ui patterns, its own flows, its own quirks. a model trained specifically on whatsapp knows exactly where the send button is, how to navigate chats, how to handle media. a model trained on instagram knows stories, reels, dms, the whole layout.

you don’t need one giant brain that kinda knows every app. you need a tiny brain that deeply knows one app. train it on that app’s screens, actions, and flows. 50M params. under 100MB. runs instantly on device.

swap models based on which app is open. that’s it.

the real bottleneck was never intelligence

it was latency. cost. battery. privacy.

micro models solve all four at once.

for mobile agents, this isn’t even a debate.