On-device AI research & products that make the assistant the only interface you need.
We build the models, runtime, and infrastructure to make on-device AI fast, capable, and accessible. On the hardware billions of people already own.
Read our research
Talk to us
The last decade shipped one app per service.
Every task earned its own icon. Rideshare. Bank. Calendar. Files. Messaging. Photos. We trained ourselves to context-switch between dozens of interfaces just to get one thing done.
Est. 8.9M mobile apps · Apple App Store + Google Play, 2026
The next decade collapses the app grid into one interface.
One instruction. Five services. Zero app launches. The assistant understands intent. Not commands. And resolves everything in a single interaction, privately, on your device.
Chase · Calendar · Files · Uber · resolved in one interaction
AI on a device needs to be fast enough that it feels less like software and more like thinking
You say what you need.
It's done. No app to open, no form to fill, no network required.
This is what the next decade looks like. Running today, on real hardware, offline. The AI capable of collapsing the app grid into one interface already exists. What Mirai is doing is making sure it runs on the device in your pocket, not just in a lab.
Technical details
1,000 t/s
Above this, the assistant stops returning text. It renders interfaces. Speed is what turns tokens into a UI
What we are building to achive 1,000 t/s on device
1
Local models
Architecture designed from device memory constraints — not compressed cloud models. Maximizes arithmetic intensity at decoding, minimizes resident set.
2
Inference engine
Built from scratch for device silicon — Metal kernels, hardware-native, not portable. The model and runtime co-evolve. No generic runtime can do this.
3
Quantization
W8A8+VQ co-designed with architecture and kernels. Hardware-accelerated weight loading. Without it, the model doesn't fit. With it, nothing is sacrificed.
4
The experience
Tokens become interfaces, forms, and resolved workflows. This is the product. Everything above is what makes it possible.
–
Silicon
Apple M-series. Already shipped — 1.4B devices worldwide. The silicon is done. The software is the open problem.
Technical details
How we build it