ludion browser-inference reality, measured

The live, CC0 dataset of in-browser LLM inference measured on real devices. Every bar below is built from people who ran the bench and shared the result. source on GitHub.

Measure your own device →

— submissions — device classes seen —

Device-class reality

Did each class actually complete a local run? Green = completed locally, red = server-routed or failed.

Routing decisions

Which rule decided where each device ran. Share of all decisions.

Speed distribution

Median local decode throughput per class. Hidden below 3 devices to protect contributors.

Failure modes

How submissions ended, as a share of all submissions.

loading…