NVIDIA's own frontier model tops US open-weights charts — and it's giving it away
What's the deal? NVIDIA runs its own frontier AI lab, and its Nemotron 3 Ultra became the top US open-weights model when it shipped a few weeks ago. Bryan Catanzaro, the company's vice president of applied deep learning research, leads the effort with hundreds of AI researchers behind him.
The chip company few realised was a model maker. NVIDIA is best known for making the chips that frontier labs train on. Now it builds and open sources foundation models too — the Nemotron family — despite selling the very compute those models consume.
Why build it? Catanzaro gives Nemotron two jobs. The first, he says, is existential: "to make sure that Nvidia continues to exist so that we can continue delivering meaningful acceleration in an era where Moore's law has died."
The logic: With transistor doubling over, NVIDIA's speed gains now come from co-designing chips, networks, compilers, and algorithms together. You cannot design silicon for AI workloads without understanding them like someone who builds the models — so Nemotron is that in-house knowledge made concrete.
Efficiency as architecture. Nemotron Ultra is a 550-billion-parameter model with 55 billion active, pre-trained in a 4-bit format, with a tenth of the model firing per token. It stacks a mostly state-space hybrid, mixture of experts served by NVL72's 72-GPU memory fabric, latent MoE for four times the experts at the same cost, and multi-token prediction for up to a 4x speedup.
The counter-intuitive pitch. Catanzaro's incentives point the other way on nearly everything he argues. He says more compute is not the answer, that the singularity is a wrongheaded idea, and that open weights are the safer option, not the dangerous one.
The signal: Open source AI is having another moment, with a powerful model landing almost weekly — and the least expected name is the one selling the compute. When the party with the most to gain from hype is the one calibrating it down, the argument is worth reading closely.
Read more: x.com
Image credit: Ch'enMeng
