hjLabs AI Playground
Starting the on-device runtime…
Everything people ask about running open models locally in a browser: what hardware you need, how big the downloads are, where the weights are stored, how to delete them, whether it works offline, what happens to your data, and where the real limits are. The short version is that a modern browser with WebGPU can run models up to a few billion parameters on your own GPU, permanently free, with nothing transmitted — and that anything beyond that ceiling still needs a server.
browser ai faqwebgpu requirementstransformers.js cachedelete cached models browseris browser ai privateoffline ai browserhow much vram browser llmclear model cache
In the browser Cache Storage API under the key "transformers-cache", keyed by the full Hugging Face resolve URL. You can inspect it in DevTools under Application → Cache Storage, and the site has a storage page that lists cached models and lets you delete them individually.
Call navigator.storage.persist(), which the storage page does for you. Without persistence, Cache Storage is best-effort and the browser may evict it under disk pressure.
Only what you download. A typical quota on a desktop browser is several gigabytes; the storage page shows navigator.storage.estimate() so you always know your usage against the quota.
No. The model weights are downloaded from Hugging Face to your browser once, and every inference after that runs on your own GPU through WebGPU. Your text, images and audio are never sent anywhere — hjLabs.in has no inference server and no way to see your input.
Yes, once a model is cached. Weights are stored in the Cache Storage API under the key "transformers-cache", so a reload with no network still runs the model. Call "Make storage persistent" first — browsers may evict non-persistent caches under disk pressure.
Small models do. WebGPU is available in recent mobile Chrome and Safari, but memory is tight — stay under about 200 MB of weights and expect the tab to be reclaimed if you switch away.