diskformer chat
A mixture-of-experts model larger than the GPU memory you give it, answering in this tab. The model downloads once into this browser's private storage (OPFS). The GPU holds the dense layers and a small cache of experts; the other experts are read from disk as each token needs them.
Chromium with WebGPU. Nothing leaves this tab except the model download from Hugging Face.
This page is inside a frame, where browser storage can be partitioned or limited. Open it in its own tab to keep the model in first-party storage.
Pick a model and a budget. The first load downloads it; later loads take seconds.
–model in this browser's storage
–GPU memory used
–experts cached on the GPU
–decode speed