Nvidia Releases Multiple Open-Source AI Models: Nemotron 3 Ultra Boosts Throughput Efficiency by 5x

At the 2026 GTC conference in San Jose, California, USA, Nvidia announced a full expansion of its open-source model family, covering agentic AI, physical AI, and medical life sciences.

In the agentic AI field, the Nemotron 3 series multi-modal models were launched: Nemotron 3 Ultra, based on the Blackwell architecture, boosts throughput efficiency by 5x for code assistance and complex workflows; Omni integrates audio, visual, and language capabilities; VoiceChat supports real-time voice conversations. Companies like CrowdStrike have deployed these models.

For physical AI, Cosmos3 will unify synthetic world generation and action simulation; Isaac GR00T N1.7 (humanoid robots) and Alpamayo1.5 (autonomous driving) are ready for commercial deployment, with the next-gen GR00T N2 set to double task success rates. In healthcare, BioNeMo added Proteina-Complexa to accelerate drug discovery, the AlphaFold database gained 1.7 million high-confidence protein complex predictions, and the nvQSP engine is 77x faster than traditional CPUs.

All models and datasets are available on platforms like GitHub, and developers can deploy them easily via NVIDIA NIM.