One model, several behaviours: LoRA adapters switched at run time

A mixture-of-experts model answers in this tab, with its experts read from disk as each token needs them. A LoRA adapter (a few to a few hundred MB, trained by someone for one job) sits on the GPU beside the model and is switched on or off between replies: no reload, no second download of the model.

Chromium with WebGPU and about 6 GB of free GPU memory. Nothing leaves this tab except the downloads from Hugging Face. The adapters are other people's work, linked to their own pages.

Model

Pick a model. The first load downloads it once into this browser's storage.

Adapters for this model

or a file

Load the model above first; then each adapter downloads once and stays on the GPU until you change model. Adapters must be llama.cpp LoRA GGUFs (F16, F32 or BF16) for this exact model. Adapters on the routed experts are refused.

Greedy decoding, so the same question gets the same answer. The engine adds scale · B(A·x) after each adapted matrix, as llama.cpp's --lora does; diskformer gates every engine against llama-server --lora. Engines: diskformer.js. Model terms: LFM Open License v1.0 (Liquid AI), Gemma (ai.google.dev/gemma/terms), Qwen3.6 (Apache-2.0).