Nvidia steps into AI model making with Nemotron 3

Nvidia is moving beyond selling AI chips by releasing Nemotron 3, a family of open models plus the training data and tools to customize them. The lineup includes Nano (30B parameters), Super (100B), and Ultra (500B). Nvidia says benchmark scores place the models among the best downloadable options. The company also shipped a hybrid latent mixture‑of‑experts design for agent building and libraries for reinforcement learning. The push arrives as OpenAI, Google, and Anthropic develop their own chips and as US firms share less about research, while Chinese companies release powerful open models more frequently. OpenRouter data shows open models handled about a third of tokens in 2025.