AI Model Playground
AI Model Playground runs Hugging Face models for 24 text, vision and audio tasks, including typed decisions, inside your browser with transformers.js; your input never leaves your device.
Starting the playground…
At a glance
- Accepts
- Text, images or audio clips, depending on the task
- Outputs
- Labels, generated text, highlighted entities, boxes drawn on your image, cut-out or depth images, speech, or sentence similarity
- Where it runs
- Entirely in your browser, on your device
- What leaves your device
- Nothing of yours - the model's weights are downloaded from Hugging Face on first use, but your input stays in the tab
- Limits
- Public transformers.js models only; model size is bounded by your device's memory, and gated models (Llama, Gemma) aren't supported
- Account
- Not required - there is no sign-up
- Price
- Free, with no watermark and no usage cap
The AI Model Playground runs machine-learning models from Hugging Face inside your browser tab. Pick a task, pick a model, give it some text, a picture or a sound clip, and see what comes back - with nothing you type or upload ever leaving your device. It covers 24 tasks across text, vision and audio, each with its own input and a result view made for it: bars for classifications, boxes drawn on your photo for object detection, a playable clip for speech, a similarity grid for embeddings.
How to use it
- Pick a task - what you want the model to do.
- Pick a model from the suggestions, or type any transformers.js model id from Hugging Face.
- Check the download size, and choose WebGPU if your browser supports it (see GPU & CPU Info).
- Give it an input and press Run.
Frequently asked questions
Is my text, image or audio sent to a server?
No. The model runs inside your browser tab with transformers.js, and your input never leaves your device. The one thing that does travel is the model itself: its weights are downloaded from Hugging Face the first time you use it, then cached by your browser.
Why does the first run take so long?
The first run downloads the model - anywhere from a few megabytes to several hundred. The button shows the exact size before you start. After that it's cached in your browser, so running it again, even after closing the tab, starts almost instantly.
What do fp32, q8, q4 and q4f16 mean?
They're how compressed the model's numbers are. fp32 is full precision and the largest download; q8 and q4 store weights in 8 and 4 bits. Counter-intuitively q4 is often bigger than q8, because only some layers get quantised - which is why this page measures each option's real download size instead of guessing.
Should I pick WebGPU or CPU?
WebGPU runs the model on your graphics card and is usually several times faster, especially for text generation and larger vision models. CPU works everywhere. If WebGPU is greyed out or fails, the GPU & CPU Info tool will tell you why.
Can I use any model from Hugging Face?
Any public model with ONNX weights made for transformers.js - type its id into the model box. Gated models such as Llama and Gemma need a Hugging Face account token, which the playground doesn't use, so they're left out of the suggestions.
Why is the text generation answer wrong?
The models small enough to download into a browser tab - 135 to 600 million parameters - are far weaker than server-hosted assistants, and they state wrong things confidently. They're useful for trying the technology, not as a source of facts.