Run AI models in your browser on your own GPU
hjLabs AI Playground runs Hugging Face models directly inside your browser, on your
computer's own GPU, using WebGPU. There is nothing to install, no account to create and no
API key to paste. A model is downloaded once from Hugging Face into the browser's cache and
from then on it runs locally — your prompts, audio and images never leave the machine,
and there is no per-token bill because there is no server doing the work.
Twenty-one task surfaces are available: chat with a local language model such as SmolLM2 or
Qwen3; transcribe a recording or your live microphone with Whisper; synthesise speech with
Kokoro 82M; remove the background from an image with RMBG 1.4; run object detection,
zero-shot detection, depth estimation, image captioning, OCR and image classification; and
on the text side, named entity recognition, sentiment analysis, zero-shot classification,
summarisation, translation, fill-mask, question answering and sentence embeddings for local
semantic search.
You need a browser with WebGPU — Chrome or Edge 113 and later, or Safari 26 and later.
Where WebGPU is missing the playground falls back to WebAssembly, which is slower but still
entirely local. Practical model sizes in the browser run from roughly 100 million to a few
billion parameters; the playground reports your adapter's real limits, including whether it
supports 16-bit shaders, before you download anything.
This page needs JavaScript to run the models. Enable it and open
ai.hjlabs.in, or browse the other developer tools at
hjLabs.in.