Nvidia released Nemotron 3.5 Lightning on August 11, an open-weight, mixture-of-experts language model with 30 billion total parameters but only 3 billion active per token, distilled from the company's larger Nemotron 3 Ultra model. The release also includes NeMo Switchyard, an open-source library that routes individual agent requests to whichever model in a fleet can handle them most efficiently, rather than sending every call to a single large model.

Built to run cheap, and fast, on one GPU

Nemotron 3.5 Lightning's sparse mixture-of-experts design activates a small fraction of its total parameters per token, letting it run on a single consumer GPU — including Nvidia's own RTX desktop cards and DGX Spark systems — rather than requiring a data-center-class setup. Nvidia says the model produces output up to four times faster than comparable dense models and completes agentic tasks roughly 30% faster than other open models in its size class, positioning it specifically for high-volume, latency-sensitive agent workloads such as tool calling and multi-step task execution rather than general chat.

Weights, training data, and licensing all public

Nvidia is releasing the model weights free for commercial use via Hugging Face and its own build platform, with no request or licensing fee required. Notably, Nvidia also published the training data and techniques behind the model, subject to underlying licensing constraints on that data — a step beyond simply open-weighting a model, which the company is framing as "truly open source" in contrast to labs that release weights without disclosing how a model was trained.

Why it matters

The release lands one day after Meta's Muse Glimmer launch and amid a broader push by U.S. labs to compete with Chinese open-weight models from Moonshot, Alibaba, and DeepSeek, which have led the open-weight leaderboard through much of 2026. Nvidia's dual move — a compact, agent-optimized model plus a routing layer to mix models of different sizes inside one agent pipeline — is a bet that the next phase of agentic AI competition is less about a single flagship model and more about the surrounding infrastructure for running many specialized models cheaply and quickly. It also reinforces Nvidia's shift from purely a chip supplier to an active participant in the open-model ecosystem it sells hardware into, a positioning that puts pressure on other infrastructure players to ship comparable open tooling of their own.