For the 10th episode of my series in which I test local AI models that fit on a 16GB MacBook Pro, I chose something different: a model that runs in your browser.

Most of the models I’ve tested so far need something like MLX Serve or a local runtime. Well, Liquid AI’s LFM3-Omni-2.2B runs entirely in the browser. You open a URL, the browser downloads roughly 1.6 GB the first time, and after that the model is just there.

Here’s the full video:

It’s a 2.2B parameter omni model, and the “omni” means it can take text, images, or video and reply with text or speech.

Adjusting The Tests for the Model

I started with simple questions. “What is a dolphin?”, my standard one, came back cleanly, both as text and as speech. Then I skipped coding and started a conversation. “Who are you?” returned a short description of the model and what it can handle. It also answered questions about my Assess-Decide-Do framework with surprising clarity — very useful for a quick lookup.

A few practical notes. This is not a coding model. At 2.2B it is too small for anything that needs real tool use or multi-step code work. What it is good for is the low-friction daily assistant chores: summarizing a document, looking at a screenshot or short video, or answering a question out loud so it feels closer to talking to a junior assistant than to a chat window.

The voice output changes a lot. Once it is speaking back, the interaction is lighter than typing into a local server. On a 16 GB machine it sits comfortably because nothing else is running a model at the same time.

All in all, it is an almost surreal experience.

The browser port was done by Xenova using custom WebGPU kernels.

• Heres’ the demo you can try on Hugging Face: https://huggingface.co/spaces/LiquidAI/LFM3-Omni-2B-WebGPU
• (Xenova’s X post is here: https://x.com/xenovacom/status/2116029909462323589).
• The base model is here: https://huggingface.co/LiquidAI/LFM3-Omni-2B.

If you like this kind of content, like, share and subscribe to my YouTube channel!