AI Performance in the Browser

On-device AI, which runs an AI model on the user's own device without sending a request to a server, is becoming more common. Running it in the browser has the advantages that nothing needs to be installed and what you type does not leave your device. In exchange, you have to download the model files, and the memory and power used for calculation are not trivial. The AI inference in Web Bench is an optional test for gauging this situation. It generates text from fixed test sentences and feeds a fixed image to the model to measure speed, and it does not read personal documents or photos. Also note that it is not an item that is always included in the standard and quick tests. It is meant only to gauge how well your device can handle AI work in the browser. The technology for running AI in the browser is developing quickly and support can widen within a few months, so if your device did not work before, it is worth updating the browser and trying again.

AI is never downloaded automatically. Only when you turn on the option to consent to the model download are the needed files fetched from the public Hugging Face CDN and jsDelivr, and the default is off. Choose after checking the download size and the memory and power use. The downloaded files may be kept in the browser cache, and you can clear them in your site data settings. If you do not consent, the AI tests are not run and are also left out of the total score. Therefore results with AI and results without it are different configurations, and they should not be mixed and compared as if they were on the same leaderboard. When you share a result, also note whether AI was included. It is also better to consider your mobile data usage and battery state at the time of download.

The execution path depends on the device. If WebGPU is available, the calculation runs on the GPU, and if not, it runs on the WASM path instead and this fact is shown separately in the result. WASM calculates on the CPU, so it can be much slower even on the same device, and this is not a malfunction but a difference in execution path. Speed and memory use also change with the precision of the model. The value is a processing speed, so higher is better, but it can change a lot if the device gets hot or memory runs short. Read it as a range repeated under the same conditions rather than a single one-off number, and accept it as a reference value. Even after the measurement ends, the result is sent to a server only with consent to join the rankings. To compare fairly you need the same model and the same input, so compare only results that share the same path and precision. You can change the selected settings at any time, and if you do not want to download anything, you can still measure the remaining tests thoroughly without AI.