Everything here is local.
Weights download once (cached in your browser), then the model generates on your GPU - offline-capable, private by construction.
Try "Explain WebGPU in two sentences" - or flip on JSON mode and ask for "the 5 largest cities in Turkey as an array of {name, population} objects" - the output is structurally guaranteed to parse.
Or flip on Tools and ask "What is 12.5% of 4,832? Use the calculator." - the model calls a calculator that runs inside this page, and the call is grammar-enforced so it can never be malformed. Every reply also shows a confidence score computed from the model's own token probabilities.
Conversations survive reloads: close the tab, come back, load the same model - the chat continues instantly, because the model's working memory (the KV cache) is saved and restored, not re-computed.
Weights download once (cached in your browser), then the model generates on your GPU - offline-capable, private by construction.
Try "Explain WebGPU in two sentences" - or flip on JSON mode and ask for "the 5 largest cities in Turkey as an array of {name, population} objects" - the output is structurally guaranteed to parse.
Or flip on Tools and ask "What is 12.5% of 4,832? Use the calculator." - the model calls a calculator that runs inside this page, and the call is grammar-enforced so it can never be malformed. Every reply also shows a confidence score computed from the model's own token probabilities.
Conversations survive reloads: close the tab, come back, load the same model - the chat continues instantly, because the model's working memory (the KV cache) is saved and restored, not re-computed.