标签：#aiampcomputing

NVIDIA shipped a 30-billion-parameter model that can see, hear, and talk — and gave the weights away. The catch: the seeing and hearing parts didn’t run on a Mac. So I spent an afternoon writing the missing piece. Here is the whole thing in thirty-five seconds — the model reading a real cart off my own store, on the laptop, with nothing leaving it. Thirty-five seconds: the model reads a real cart off my own store, on the laptop, with nothing leaving it. The elapsed counter is the real measured latency. Every couple of weeks I go looking for whatever new open-weight model just dropped, pull it onto my laptop, and see what it can actually do. Most of the time it’s a coding model, I run it against the one I already use, and my current favorite wins again. That’s a fine result. It’s just not much of a story. This time I found something different. What Nemotron Omni actually is NVIDIA released Nemotron-3-Nano-Omni-30B-A3B — a “tri-modal” model, which is a fancy way of saying one brain with eyes and ears attached. You can hand it a picture, a sound file, or a video, and talk to it about what it saw or heard. It’s 30 billion parameters total, but only about 3 billion of them fire for any given word, which is why something this capable can run on a laptop at all. The weights are public. Someone had already done the hard, unglamorous work of shrinking it down to a 4-bit MLX version that fits on Apple Silicon — that’s yayr over at mlx-community, and this project doesn’t exist without that upload. About 19 GB on disk. Ready to go. Except for one line, buried in the model card: The text backbone loads with standard MLX nemotron_h tooling. The vision and audio towers require a multimodal runtime that implements the C-RADIO ViT-H and Parakeet Conformer forward passes (e.g. the Evorix on-device engine). Translated: the brain works on a Mac. The eyes and ears don’t. The weights for them are right there in the file — all the knowledge, sitting on your disk — but nothing I could find