As always, the demo is open source (which you can find under the "Files" tab), so I'm excited to see how the community builds upon this! 🚀

🔗 Link to demo: huggingface.co/spaces/Liqui...

That's right, we're running Mistral's new Voxtral-Mini-3B model 100% locally in-browser on WebGPU, powered by Transformers.js and ONNX Runtime Web! 🔥

Try it out yourself! 👇
huggingface.co/spaces/webml...

Model: huggingface.co/lazy-guy12/c...

Online demo: lazy-guy.github.io/chess-llama/

The most difficult part was getting the model running in the first place, but the next steps are simple:
✂️ Implement sentence splitting, enabling streamed responses
🌍 Multilingual support (only phonemization left)

Who wants to help? 🤗
huggingface.co/spaces/webml...

The model is also extremely resilient to quantization. The smallest variant is only 86 MB in size (down from the original 326 MB), with no noticeable difference in audio quality! 🤯

Link to models/samples: huggingface.co/onnx-communi...

You can get started in just a few lines of code! 🧑‍💻

Huge kudos to the Kokoro TTS community, especially taylorchu for the ONNX exports and Hexgrad for the amazing project! None of this would be possible without you all! 🤗

Try it out yourself: huggingface.co/spaces/webml...

For the AI builders out there: imagine what could be achieved with a browser extension that (1) uses a powerful reasoning LLM, (2) runs 100% locally & privately, and (3) can directly access/manipulate the DOM! 👀

💻 Source code: github.com/huggingface/...
🔗 Online demo: huggingface.co/spaces/webml...

This project was greatly inspired by Brendan Bycroft's amazing LLM Visualization tool – check it out if you haven't already! Also, thanks to Niels Rogge for adding DINOv2 w/ Registers to transformers! 🤗

Source code: github.com/huggingface/...

Online demo: huggingface.co/spaces/webml...

Another interesting thing to see is how the attention maps become far more refined in later layers of the transformer. For example,

First layer (1) – noisy and diffuse, capturing broad general patterns.
Last layer (12) – focused and precise, highlighting specific features.

Vision Transformers work by dividing images into fixed-size patches (e.g., 14 × 14), flattening each patch into a vector and treating each as a token.

It's fascinating to see what each attention head learns to "focus on". For example, layer 11, head 1 seems to identify eyes. Spooky! 👀

The app loads a small DINOv2 model into the user's browser and runs it locally using Transformers.js! 🤗

This means you can analyze your own images for free: simply click the image to open the file dialog.

E.g., the model recognizes that long necks and fluffy ears are defining features of llamas! 🦙

Yeah, I ran into this during development, and is unfortunately a bug in Firefox:
- bugzilla.mozilla.org/show_bug.cgi...
- bugzilla.mozilla.org/show_bug.cgi...

Huge shout-out to the Useful Sensors team for such an amazing model and to Wael Yasmina for his 3D audio visualizer tutorial! 🤗

‍💻 Source code: github.com/huggingface/...
🔗 Online demo: huggingface.co/spaces/webml...

Huge shout-out to OuteAI for their amazing model (OuteTTS-0.2-500M) and for helping us bring it to the web! 🤗 Together, we released the outetts NPM package, which you can install with `npm i outetts`.

💻 Source code: github.com/huggingface/...

🔗 Demo: huggingface.co/spaces/webml...

The model is multilingual (English, Chinese, Korean & Japanese) and even supports zero-shot voice cloning! 🤯 Stay tuned for an update that will add these features to the UI!

More samples:
bsky.app/profile/reac...