Nvidia has introduced Nemotron 3, the latest generation of its open model family, aimed at helping enterprises build and scale increasingly complex agentic AI applications. Announced in mid-December 2025, the family arrives in the same three sizes as its predecessor — Nano, Super, and Ultra — and is built on a hybrid mixture-of-experts (MoE) architecture that, according to Nvidia’s announcement, targets developers deploying multi-agent systems with better transparency and efficiency.
The lineup
Nemotron 3 Nano is a 30-billion-parameter model that activates only a few billion parameters per token, designed for highly efficient operation; it is available immediately. Nemotron 3 Super is a high-accuracy reasoning model with roughly 100 billion parameters and about 10 billion active per token, positioned for multi-agent applications. Nemotron 3 Ultra is a large-scale reasoning model of roughly 500 billion parameters for the most complex workloads. Super and Ultra are slated to follow Nano in the first half of 2026. The family pairs a Mamba-Transformer hybrid design with MoE routing and supports context lengths up to one million tokens.
Alongside the models, Nvidia released around three trillion tokens of new Nemotron pre-training, post-training, and reinforcement-learning datasets, plus the open-source NeMo Gym and NeMo RL libraries, which provide a training environment and post-training foundation for Nemotron models.
Iteration, not revolution
Analyst reaction has been measured. Lian Jye Su of Omdia describes Nemotron 3 as impressive but not revolutionary — an iteration of the previous Nemotron generation that gets smarter with each release. In his view, the models’ popularity in the developer community gives Nvidia good reason to keep investing in the line.
The multi-agent scaling problem
With this release Nvidia is squarely targeting one of the AI market’s current challenges: scaling multi-agent systems. Nvidia says Nemotron 3 Nano delivers the highest token throughput per second for multi-agent systems — several times its predecessor — and its long context window lets a model retain more information across long, multi-step tasks.
Multi-agent systems remain difficult to implement. As Su describes it, the complexity lies in a model’s ability to decompose work: breaking a problem into tasks, understanding each task’s requirements, and assigning the right task to the right place. An emerging pattern uses smaller models to execute individual tasks while a larger model supervises the overall instruction — an architecture Nemotron’s size range maps onto naturally.
The enterprise security question
Open models carry a specific adoption hurdle: enterprise-grade security and governance support. Su notes that previous Nemotron generations shipped with guardrails built into the models that many enterprises considered inadequate, leading most enterprise users to build an additional security layer on top — which itself requires in-house expertise. His blunt assessment: organizations without strong internal capability in systems engineering for multi-agent setups are unlikely to adopt an open model blindly, and may be better served by managed offerings from vendors such as Anthropic or OpenAI.
Limitations and what to watch
Several caveats frame this release. Nvidia’s throughput and efficiency figures are vendor benchmarks, and independent evaluations of Nemotron 3 — particularly of Super and Ultra, which were not yet generally available at announcement — will take time to accumulate; the team’s technical report on arXiv provides the methodological detail. The open-weights license terms and the practical cost of self-hosting large MoE models are decisive factors for enterprise adoption that headline parameter counts do not capture. Worth watching: whether Super and Ultra ship on schedule in early 2026, how the models score on independent agentic benchmarks, and whether Nvidia strengthens built-in guardrails in response to the enterprise criticism. Related reading on this site: LongCat-2.0 and open-source models for small business and frontline lessons on building enterprise AI agents.
