LINUS AI · local inference
What can your machine actually run?
Not a demo of a toy. This measures the machine you are on, tells you the largest model it can run locally and roughly how fast, and then lets you verify the engine right here — before you install anything.
Your machine can run
Throughput figures are measured on real hardware in our own fleet with this engine, not modelled. Your machine will differ; the ranking will not.
Prove it, here, now
The same engine compiled to WebAssembly. It downloads a model once and runs it in this tab — no install, no account, and nothing you type leaves the machine.
Why the browser is the slow way — and why that is the point
In a tab the engine gives up the two things it is built on: memory-mapping the model (natively the file is the arena, so RAM cost equals file size and the OS pages it in) and multi-core SIMD. It runs single-threaded on a copy in memory. Measured on the same machine, same model, same engine:
| model | in a browser tab | installed |
|---|---|---|
| 360M | 3.4 tok/s | 28.6 tok/s |
| 1B | 1.1 tok/s | 10.2 tok/s |
So: the tab proves the engine runs on your machine with nothing installed and nothing sent anywhere. Installing it is what makes local AI a working tool rather than a curiosity — roughly 9× faster, and it scales to models a tab cannot hold at all.
Get the build for this machine
The engine is Apache-2.0 and free. There is one binary per machine — this picks yours, and the list is read from the store itself, so nothing is offered here that is not actually published.