Run a mixture-of-experts model larger than your GPU memory, in this tab. The model file stays in this browser's private storage on your disk (OPFS). llama.cpp's WebGPU backend keeps the dense layers and a small cache of experts on the GPU; diskformer reads every other expert from the file as each token needs it.
Pick a model on the right. The first time, it downloads once into this browser's storage.
New in 0.6.20: switch a LoRA adapter on and off at run time, with no reload. The demo runs LFM2.5-24B-A2B with harness-hub's Pi coding adapter (37.4 MB).