NVIDIA Unveils Open-Source AI Model: Nemotron 3 Ultra Achieves 5-Fold Increase in Throughput Efficiency

K-LinePoet · 2026-03-23T13:09:51+00:00

Nvidia announced at its 2026 GTC conference an expansion of its open-source model family, introducing the Nemotron3 series of multimodal models to enhance capabilities in agents, physics, and medical AI. The company also launched new models to accelerate autonomous systems and drug discovery, while enhancing GPU parallel computing performance, with open-source resources now available across multiple platforms.

K-LinePoet

2026-03-23 13:09:51

Abstract generation in progress

NVIDIA GTC 2026 Conference Highlights

On March 17, at the GTC conference held in San Jose, California, NVIDIA announced a comprehensive expansion of its open-source model family to promote the next wave of AI in agents, physical systems, and healthcare.

In the field of agents, NVIDIA introduced the Nemotron3 series of fully understanding multimodal models. Among them, Nemotron3 Ultra, based on the Blackwell architecture, offers a fivefold increase in throughput efficiency, designed for code assistance and complex workflows.

Nemotron3 Omni precisely integrates audio, visual, and speech multimodal capabilities; Nemotron3 VoiceChat combines speech recognition, large models, and speech synthesis to support real-time conversations.

The accompanying security models further ensure the reliability of content output, with leading companies like CrowdStrike and ServiceNow already deploying them in enterprise applications.

In the physical AI domain, NVIDIA released several foundational models to accelerate autonomous system development. The upcoming Cosmos3 is the first unified world synthesis and motion simulation foundational model; Isaac GR00T N1.7 and Alpamayo1.5 are visual language action (VLA) reasoning models designed for humanoid robots and autonomous driving, now ready for commercial deployment.

Additionally, CEO Jensen Huang previewed the next-generation robot model GR00T N2, which doubles the task success rate in new environments and is expected to launch by the end of the year.

Major breakthroughs in medical life sciences research have also been achieved. According to a blog post cited by IT Home, the BioNeMo platform added the generative model Proteina-Complexa, significantly accelerating protein drug discovery.

At the same time, NVIDIA partnered with organizations like Google DeepMind to add 1.7 million high-confidence protein complex predictions to the AlphaFold database. Moreover, the new GPU-accelerated simulation engine nvQSP offers speeds up to 77 times faster than traditional CPUs, enabling scientists to analyze vast treatment options in a very short time.

These open-source models, datasets, and frameworks are now available on platforms like GitHub and Hugging Face. Developers can also utilize NVIDIA NIM microservices to achieve secure, scalable, and convenient deployment across edge and cloud infrastructure.

NVIDIA GTC 2026 Conference Special

View Original

This page may contain third-party content, which is provided for information purposes only (not representations/warranties) and should not be considered as an endorsement of its views by Gate, nor as financial or professional advice. See Disclaimer for details.