Nvidia on Monday announced the Nemotron 3 family of openly released AI models, training datasets, and engineering libraries.
This marks an aggressive push into open-source AI development.
The move signals Nvidia’s intent to dominate not just the hardware layer of artificial intelligence, but also the software and model layers.
The development comes amid companies worldwide seeking domestic, auditable alternatives to closed or foreign AI systems.
The release bundles model weights, a synthetic pretraining corpus of nearly 10 trillion tokens, and detailed training recipes under an open license.
It allows developers and enterprises to inspect, customize, and deploy Nemotron models on their own infrastructure.
The strategic calculation is transparent, as open-source AI proliferates globally.
With the government agencies in the US demanding transparency, Nvidia is positioning itself as the trusted domestic supplier, while fortifying its ecosystem of developers.
What Nvidia released: Models, data, and technical claims
The Nemotron 3 family consists of three models in increasing sizes: Nano (30 billion parameters with 3 billion active), Super (100 billion with 10 billion active), and Ultra (500 billion with 50 billion active).
Only Nemotron 3 Nano ships immediately; Super and Ultra arrive in the first half of 2026.
Nvidia’s marquee claim is efficiency. Nemotron 3 Nano delivers four times the throughput of its Nemotron 2 predecessor and reduces reasoning token generation by up to 60%.
The models employ a hybrid latent mixture-of-experts architecture, a design that activates only the most relevant computational pathways for each task, mimicking how the human brain compartmentalizes work.
This approach has become the industry standard, with the top 10 most intelligent open-source models now using MoE, according to independent benchmarking data.
It must be noted that the amount of information a model can hold in memory expands to one million tokens for Nano, seven times broader than its predecessor.
That matters for long-form documents, code repositories, and complex multi-step reasoning.
Super and Ultra leverage Nvidia’s 4-bit NVFP4 training format on its Blackwell hardware, slashing memory requirements and training time without sacrificing accuracy.
All model weights, training corpora, and detailed recipes are available on GitHub and Hugging Face under the NVIDIA Open Model License.
Developers also gain access to NeMo Gym, NeMo RL, and NeMo Evaluator, the open-source libraries for training, reinforcement learning, and safety validation.
Why open release matter?
The open release directly responds to mounting enterprise demand for model transparency.
“Many of our enterprise customers cannot deploy certain models or build their business on models with opaque source codes,” said Kari Briski, Nvidia’s VP of generative AI software.
Regulated industries like healthcare, finance, and defense require auditable, on-premises alternatives to proprietary systems controlled by foreign entities.
Nvidia’s move gains force as Meta retreats from open-source. Llama’s growth stalled after April’s lukewarm Llama 4 launch, ceding ground to rival open models.
Meta has withheld training datasets even from close partners like Nvidia, limiting community-driven improvements.
By contrast, Nvidia publishes everything: weights, recipes, and datasets. That transparency could attract enterprise customers and government contracts wary of opaque dependencies.
Geopolitically, US tariffs and export restrictions on Chinese AI amplify Nvidia’s advantage.
The post Nvidia unveils Nemotron 3: why is NVDA making its latest AI models open source? appeared first on Invezz