NVIDIA's Nemotron3 Super/Ultra models reportedly using groundbreaking FP4 pre-training
NVIDIA's secret weapon for next-gen AI models could change the hardware game.
Deep Dive
In a recent interview, NVIDIA's VP revealed the company is pre-training its upcoming Nemotron-3 Super and Ultra models using FP4 precision, a technique not yet done publicly. The models are slated for a broad H1 2026 release. The VP also stated NVIDIA's open-source model strategy is to drive GPU sales and described the company's culture as decentralized, calling employees "volunteers" who self-organize on projects like Nemotron.
Why It Matters
FP4 training could drastically reduce AI compute costs and solidify NVIDIA's hardware dominance for years.