NVIDIA Corp. on Monday introduced Nemotron 3, a family of open-source artificial intelligence (AI) models that position the chipmaker to capture the growing market for specialized AI agent systems amid a flurry of offerings from China.
The new models — Nano, Super, and Ultra — are built on what NVIDIA calls a breakthrough hybrid latent mixture-of-experts architecture designed to address key challenges developers face when building multi-agent AI systems, including communication overhead, context drift, and high inference costs.
“Open innovation is the foundation of AI progress,” NVIDIA CEO Jensen Huang said in a statement. “With Nemotron, we’re transforming advanced AI into an open platform that gives developers the transparency and efficiency they need to build agentic systems at scale.”
The Nemotron 3 Nano model, available immediately, features 30 billion parameters with 3 billion active and delivers up to four times higher token throughput compared to its predecessor while reducing reasoning-token generation by up to 60%. The model includes a one-million-token context window, allowing it to maintain accuracy across extended, multistep tasks.
Artificial Analysis, an independent AI benchmarking organization, ranked Nemotron 3 Nano as the most open and efficient model in its size category with leading accuracy metrics.
NVIDIA is entering the open-source model space as Chinese companies like DeepSeek, Moonshot AI, and Alibaba Group Holdings gain traction in the tech industry. Airbnb, for instance, has publicly adopted Alibaba’s Qwen open-source model. Meanwhile, according to CNBC and Bloomberg, Meta Platforms Inc. may be moving away from open-source toward closed-source models, which would position NVIDIA among the leading U.S. providers of open-source AI offerings.
Major technology firms are already integrating Nemotron models into their operations. Accenture, Cadence Design Systems Inc., CrowdStrike Holdings Inc., Deloitte, Oracle Cloud Infrastructure, Palantir Technologies Inc., Perplexity AI, ServiceNow Inc., Siemens, and Zoom Communications Inc. are implementing the technology across manufacturing, cybersecurity, software development, and communications sectors.
ServiceNow CEO Bill McDermott emphasized the partnership’s potential, stating the combination of ServiceNow’s workflow automation with Nemotron 3 would “define the standard with unmatched efficiency, speed and accuracy.”
Perplexity CEO Aravind Srinivas highlighted the model’s role in optimizing workload distribution, noting that their agent router can direct tasks to fine-tuned open models like Nemotron 3 Ultra or leverage proprietary models when specific capabilities are needed.
NVIDIA also released three trillion tokens of training datasets and new reinforcement learning libraries through its NeMo platform, including the NeMo Gym and NeMo RL open-source tools available on GitHub and Hugging Face.
The Nemotron 3 Nano is currently accessible through Hugging Face and multiple inference service providers, with availability on Amazon Bedrock and Google Cloud expected soon. The larger Super and Ultra models are scheduled for release in the first half of 2026.
