Which models can you actually run in the browser, on the visitor's own GPU, with no server and no API bill? This tool detects your browser's WebGPU support live, then shows the chat and embedding models that fit a GPU budget you set, with honest quantized download sizes. It runs entirely in your tab. Background: running your own model cross-browser and the Browser AI Readiness Probe.
WebLLM runs chat models on the GPU through WebGPU, so a graphics chip is required for the chat models. Pick your GPU memory and system RAM; the tool checks each quantized model against a budget with headroom for the runtime and context. Embedding and small vision models run through Transformers.js and fall back to the CPU, so they fit almost anywhere.
engineering judgment Browsers do not expose exact VRAM, so pick the closest tier. Fit is a rule of thumb (quantized weights plus roughly 30 to 40% headroom for the runtime and context), not a guarantee. Real usage moves with context length and what else is on the GPU.
Pick a small chat model, load it into your browser through WebLLM, and run it on your GPU. The first load downloads the model (sizes shown) and caches it; nothing you type leaves your device. It needs WebGPU and downloads real weight, so it is opt-in. You can switch models afterward.
experimental A live demo of the on-device path from the bring-your-own-model post. If your browser lacks WebGPU or the load is blocked, you get a clear message instead of a broken box. Note: tiny models like this often misidentify themselves (it may claim to be “ChatGPT”) — that is a hallucination, not a network call. The site’s policy does not permit it to reach any chat API; the model and every reply stay on your device. Prove it by turning off your wi-fi and asking again.
The advice "just run the model in the browser" skips the part that actually stops people: the model has to fit in the visitor's GPU memory, and the file has to download first. A model that runs beautifully on a 16 GB GPU simply will not load on an integrated one. This tool turns that into a straight answer for a budget you choose, and detects whether your current browser even has WebGPU, so you plan for the hardware your visitors actually have instead of the machine you built it on. Sizes are approximate MLC q4f16 figures; treat them as planning numbers and verify against the WebLLM model list.