NVIDIA Releases Nemotron 3, Open Models Tuned for AI Agents

NVIDIA unveiled the Nemotron 3 family of open models at its developer day on Tuesday, positioning the release as a foundation for agentic applications rather than another general-purpose chatbot. The family comes in three sizes, Nano, Super and Ultra, and is paired with companion model lines for speech, retrieval and safety.

The smallest model carries most of the early news. Nemotron 3 Nano delivers four times the throughput of its predecessor while keeping accuracy competitive with proprietary alternatives, NVIDIA says, thanks to a hybrid mixture-of-experts architecture that routes work selectively across the model. For multi-agent workloads, where many small calls happen in parallel, raw tokens per second is the metric that matters, and Nano is built for that pattern.

A wider stack, not just a model

Alongside the core models, NVIDIA released Nemotron Speech, a leaderboard-topping family of speech recognition models designed for low-latency live captioning, and Nemotron RAG, a set of embed and rerank vision-language models that handle multilingual and multimodal retrieval. A separate Nemotron Safety line is intended to strengthen the safety properties of applications built on the family.

A naked language model is no longer enough on its own. Production agents need speech, retrieval and guardrails alongside the model itself, and shipping all of those as open releases under a single brand makes the stack easier to assemble. Whether the safety models actually do useful work in practice is another question, but the structural choice is striking on its own.

Why open matters here

NVIDIA's open releases sit awkwardly alongside its core hardware business, where it sells the GPUs every closed lab also depends on. Open weights cost the company nothing in chip demand, and arguably increase it, because every fine-tuned downstream model still has to run somewhere. The strategic logic is plain enough.

For developers, what matters is that Nemotron 3 lands in a year when frontier open weights from US labs have been scarce. Cursor's Composer 2.5 release earlier this month addressed the coding case, but a broader, hardware-vendor-backed family of agentic models is a different proposition. Super and Ultra are expected later this year, and the question will be whether the larger sizes hold their own against closed frontier models when they do arrive.