Overview
- Nvidia released the Nemotron 3 Nano model today as open source, with 100‑billion‑parameter Super and 500‑billion‑parameter Ultra versions planned for the first half of 2026.
- The company published three trillion tokens of pretraining data and 18 million post‑training samples, along with tools that let enterprises and governments test security and customize deployments.
- The new family adopts a mixture‑of‑experts architecture that activates only a subset of parameters to boost throughput and cut compute costs compared with prior Nemotron models.
- Nvidia says Nano delivers up to 4x performance over Nemotron 2 Nano, produces up to 60% fewer reasoning tokens, and supports a 1 million‑token context window.
- The launch coincides with rapid uptake of Chinese open models such as Alibaba’s Qwen at companies like Airbnb, U.S. restrictions on Chinese AI use by public entities, and reports that Meta may move toward closed releases.