News Summary
- Nvidia released Nemotron 3.5 Lightning on August 11, 2026, a 30-billion-parameter open-source model with 3 billion active parameters per token.
- The model is distilled from Nvidia’s larger Nemotron 3 Ultra and is designed for high-volume agentic AI workloads.
- Nvidia CEO Jensen Huang made his first post on X on July 24, 2026, publishing an industry letter urging Washington to avoid restricting open-weight AI models.
- The letter, titled “Open Weights and American AI Leadership,” launched with 25 signatories and grew to more than 150 organizations within days.
- Nvidia also released NeMo Switchyard, an open-source library that routes AI agent requests to the most efficient model for each task.
- The release comes amid a Washington debate over Chinese open-weight models, particularly Moonshot AI’s Kimi K3.
Nvidia has released Nemotron 3.5 Lightning, a 30-billion-parameter open-source artificial intelligence model designed to power autonomous AI agents. The announcement on August 11 marks the company’s first open-weight model release since Chief Executive Jensen Huang entered the public debate over open-source AI policy in Washington less than three weeks ago.
Nemotron 3.5 Lightning uses a mixture-of-experts architecture that activates only 3 billion parameters per token, allowing it to run on a single consumer GPU such as those in Nvidia RTX PCs or the company’s DGX Spark systems. Nvidia said the model delivers up to four times faster output speed and completes agentic tasks roughly 30 percent faster than comparable open models in its class.
The model is available free of charge for commercial use. Companies can download the weights from Hugging Face and Nvidia’s build platform without requesting permission or paying licensing fees. Nvidia also published the training data and techniques used to create the model, subject to licensing constraints.
The timing of the release is notable. On July 24, Huang published his first-ever post on X to share an open letter signed by Nvidia and more than two dozen technology companies urging U.S. policymakers not to impose broad restrictions on open-weight AI models. The letter, titled “Open Weights and American AI Leadership,” argued that downloadable model weights strengthen safety, accelerate innovation, and preserve national sovereignty.
“Open models strengthen safety and cybersecurity, accelerate innovation and diffusion, and enable sovereignty,” Huang wrote in the post. “The world needs both frontier closed models and frontier open models.”
The letter’s initial 25 signatories included Microsoft, Meta, Palantir, Hugging Face, and IBM. OpenAI and Google added their names within days, though Anthropic did not sign. By July 28, the roster had grown to more than 150 organizations, according to Microsoft’s published list of signatories.
The industry push arrived as the Trump administration weighs its stance on Chinese open-weight models. Moonshot AI released Kimi K3 on July 16, a model that ranks among the most capable open-weight systems available. White House Office of Science and Technology Policy Director Michael Kratsios accused Moonshot of developing Kimi K3 by distilling from Anthropic’s Fable 5, a technique that trains a smaller model using outputs from a more advanced one. Treasury Secretary Scott Bessent suggested the United States could sanction companies found to have improperly used distillation to extract intellectual property.
The coalition letter acknowledged that unlawful extraction from closed models raises legitimate concerns, but argued those issues should be addressed through targeted legal frameworks rather than sweeping restrictions on distillation itself. Nvidia representatives confirmed that Nemotron 3.5 Lightning was developed using distillation from the company’s larger Nemotron 3 Ultra model.
Three days after Huang’s X post, Nvidia and 36 other organizations formed the Open Secure AI Alliance to develop open technologies for securing AI agents. Members include Microsoft, Cisco, CrowdStrike, and the Linux Foundation. The alliance’s first technical contribution is NOOA, an Apache 2.0 research framework for testing and governing agent behavior.
Nvidia is not the only major technology company pressing the open-source case. Meta Chief Executive Mark Zuckerberg published a manifesto arguing for open-source AI around the same time the company released Muse Code, a terminal-based coding agent, and an open-source variant of its Muse Spark model. “Our goal should be for American open source models to be the best globally,” Zuckerberg wrote.
For Nvidia, open-source models serve a straightforward business purpose: they still require GPUs to run, and broader adoption drives chip sales. Huang made the connection explicit in an interview with Axios last month. “Free AI should be great for hardware,” he said. “Free AI should be great for chips.”
Nemotron 3.5 Lightning is built for what Nvidia calls “always-on agents”,AI programs that operate autonomously in the background, gathering context, calling tools, and working through multi-step tasks. The model supports context windows of up to 1 million tokens and can be post-trained on an organization’s own data using Nvidia’s NeMo framework.
Several companies have already customized the model for specialized workloads. CrowdStrike is using it for cybersecurity applications, Harvey with Trajectory for legal services, and CodeRabbit with Baseten for code review. Lila Sciences and Fastino Labs have also customized the model for scientific and enterprise tasks, according to Nvidia.
Alongside the model, Nvidia released NeMo Switchyard, an open-source routing library written in Rust that directs agent requests to the most capable and cost-efficient model for each step of a workflow. Nvidia said internal benchmarks showed that a Switchyard-routed system mixing open models with Anthropic’s Opus 4.8 maintained frontier-level accuracy while cutting task-completion costs to roughly one-third of running Opus alone.
The model is available in BF16 and NVFP4 formats through Hugging Face, OpenRouter, and Nvidia’s own build platform. Nvidia said it is working with partners including LangChain, Kong, and Cognition AI to integrate Switchyard into existing developer tools.