NVIDIA announced Nemotron 3, a new family of open models designed specifically for building multi-agent AI systems—positioning it as a major step in the company’s push into frontier model development. The lineup includes three sizes: Nano (30B), Super (100B), and Ultra (500B). NVIDIA says Nano is available now, while Super and Ultra are expected to arrive in 2026. According to NVIDIA, Nemotron 3 Nano outperforms similar-sized models (e.g., Qwen3-30B) on coding and instruction-following benchmarks, while also generating responses over 3x faster. NVIDIA also emphasized openness beyond weights: it is publishing training data, fine-tuning tools, and reinforcement learning environments alongside the models. The company highlighted Cursor, Perplexity, ServiceNow, and CrowdStrike as early adopters across coding, search, enterprise automation, and cybersecurity. Strategically, the release aims to strengthen the Western open-model ecosystem while keeping developers building on NVIDIA hardware as open-source adoption accelerates globally.