Amid intensifying interest in Chinese open-weight models, AI hardware vendor Nvidia has introduced its latest open source model: Nemotron 3.5 Lightning. Released on August 11, the model is designed to be small enough for enterprises to run on local devices while targeting specific tasks within larger multi-agent systems.

Nemotron 3.5 Lightning is a mixture-of-experts model — an architecture that activates only a subset of the model's total parameters for any given query, allowing it to deliver stronger performance at lower inference cost than a comparably sized dense model. It is capable of performing specialized functions such as code review, tool use, security alert monitoring, and answering billing questions, Nvidia said. The company released the model weights, code, recipes, and training data under the Linux Foundation's OpenMDW 1.1 license, making it a fully open source model.

The model is part of Nvidia's Nemotron 3 family of open models, originally released last December. It has 30 billion parameters — smaller than Nemotron 3 Ultra (550 billion) and larger than Nemotron 3 Nano (8 billion) — and is intended to help enterprises build agentic applications on local devices.

Alongside Nemotron 3.5 Lightning, Nvidia released NeMo Switchyard, an open source library for routing within popular agentic applications. Enterprises can use Switchyard to route a prompt or agentic request to the most suitable model across their own mix of open, proprietary, or Nvidia models, the company said. Model routing tools have become an increasingly important piece of enterprise AI stacks as organizations adopt multi-model architectures, assigning different tasks to different models based on cost, latency, and capability requirements.

Both releases come less than a month after Nvidia CEO Jensen Huang revealed that the AI hardware-software giant had joined a broad coalition of tech vendors defending open source, as the Trump administration weighed restrictions on open-weight models from China. In an X post, Huang said open models enable sovereignty and strengthen safety and cybersecurity.

Nemotron 3.5 Lightning is Nvidia's first open model since debate over advancements in Chinese open models intensified last month, when Chinese AI vendors Alibaba and Moonshot AI released inexpensive, powerful open-weight models. The Nvidia launch also comes a day after social media giant Meta released its latest open model.

"It's a testament to just how impactful the Chinese models have been and will be," said Bradley Shimmin, an analyst at Futurum Group. He added that many consider Alibaba Qwen 3.8 Max to be "the benchmark for local capable open source agentic development."

Supporting Nvidia's Infrastructure

For Nvidia, providing open source models is not only a competitive strategy against Chinese vendors but also a way to help sell its chips and hardware, according to Arun Chandrasekaran, an analyst at Gartner.

"Models help Nvidia sell its infrastructure better," Chandrasekaran said. "When you sell a model, you have to run the model somewhere. Running the models requires networking, inference software and data training."

"It is in Nvidia's own interest that there is a vibrant model ecosystem that's out there, and the model ecosystem is not very concentrated," Chandrasekaran continued. He noted that, for the most part, enterprises use Nvidia's models for specific tasks, on-premises deployments, or scenarios with significant sovereignty considerations.

With Nemotron 3.5 Lightning, Nvidia said enterprise users can run the model on local AI systems, including Nvidia's RTX PCs, DGX Spark, and Jetson. The emphasis on local deployment reflects a broader enterprise shift toward on-premises AI infrastructure, driven by data sovereignty regulations, privacy compliance requirements, and the need to control ongoing operational costs.

Training Data and Model Size

Nvidia's decision to release the training data alongside the model is commendable, Shimmin said. "There are very few vendors, with the exception of IBM, that have been promoting and touting this ability to say, 'if you indeed care about exposure to risk in adopting AI, you should care about the training data used to build these models,'" he said. "It all starts with the data, and it honestly ends with the data too."

The 30-billion-parameter size of Nemotron 3.5 Lightning illustrates how vendors like Nvidia are beginning to acknowledge the need to "right-size" models, Shimmin added. He said many vendors realize that the current approach of using APIs to access models on a token-per-cost basis is too expensive, and working with models locally helps enterprises justify the cost of generative AI. The release also reflects a wider industry trend in which vendors including Meta, Mistral, and Alibaba have introduced smaller, task-oriented models alongside their flagship systems, recognizing that many enterprise use cases do not require frontier-scale capabilities.

"Tokenomics over time have already proven to be a little bit hazardous for companies in terms of both not understanding their token spend and having no ability to actually gain observability or transparency in how that spend is going," Shimmin said.