NVIDIA has taken a step that directly addresses the core of the next wave of software: multi-agent systems. The company has announced Nemotron 3, a new family of open models—available in Nano, Super, and Ultra sizes—along with datasets and libraries for training and post-training, with a clear goal: making it cheaper, more transparent, and easier to deploy specialized agents.
This isn’t accidental. The market is moving away from single chatbots and embracing architectures where multiple agents divide tasks, coordinate, and correct each other. But this coordination comes at a cost: more tokens, higher latency, increased inference costs, and greater risk of “context drift” as flows extend. In this scenario, NVIDIA aims to combine two promises that don’t always go hand in hand: efficiency and openness.
A “Hybrid” MoE to Reduce Costs in Multi-Agent Environments
The technical cornerstone of this announcement is a hybrid latent mixture-of-experts (MoE) architecture. In practical terms, this means the model doesn’t activate its entire size with each token, but rather turns on a subset — the “experts” — to handle what’s needed at that moment. NVIDIA assures that this approach enables Nemotron 3 Nano to achieve up to 4× higher throughput compared to Nemotron 2 Nano, and to cut reasoning token generation by up to 60%, focusing on reducing costs in long flows and systems with many concurrent agents.
This efficiency is particularly relevant for companies building model “routers”: using a very powerful proprietary model for high-value, specific tasks, while offloading other work (summarization, extraction, classification, internal support, retrieval) to more controllable open models with predictable costs. NVIDIA emphasizes that “tokenomics”—the actual cost of operating an agent—is becoming a strategic variable.
Three Sizes: Nano Now; Super and Ultra in 2026
The family comes in three tiers:
- Nemotron 3 Nano: a model with 30 billion parameters that activates up to 3 billion per token. Positioned as the “workhorse” for efficient tasks: software debugging, summaries, assistant flows, and IR (search/retrieval) with low cost.
- Nemotron 3 Super: geared toward high-precision reasoning for multi-agent applications, with approximately 100 billion parameters and up to 10 billion active per token.
- Nemotron 3 Ultra: designed as a reasoning engine for complex tasks, with around 500 billion parameters and up to 50 billion active per token.
Availability-wise, NVIDIA sets a clear milestone: Nano is available now, while Super and Ultra are expected in the first half of 2026.
And here’s a noteworthy detail, which for some use cases may matter even more than model size: Nemotron 3 Nano boasts a context window of 1,000,000 tokens. This figure indicates suitability for long flows, operational memory, and multi-step tasks where the model must “hold” a significant history without breaking down.
“It’s Not Just a Model”: Datasets and Libraries for Training Agents
The announcement isn’t limited to open weights. NVIDIA claims to be “the first” to release a comprehensive package of models + datasets + reinforcement learning environments/libraries aimed at creating specialized agents with precision and efficiency. Specifically, they mention three trillion tokens in datasets for pretraining, post-training, and RL, in addition to a Nemotron Agentic Safety Dataset to evaluate and enhance safety in complex agent systems.
On the tooling side, the company launches NeMo Gym and NeMo RL as open-source libraries for training and post-training environments, along with NeMo Evaluator for safety and performance validation. All available on GitHub and Hugging Face, with integrations mentioned with key ecosystem players.
Ecosystem and Deployment: From Laptop to Enterprise
NVIDIA is working to ensure Nemotron 3 doesn’t stay confined to labs. Compatibility and distribution support widely used tools and runtimes such as: LM Studio, llama.cpp, SGLang, and vLLM.
They also detail availability of Nemotron 3 Nano on Hugging Face and through inference providers like Baseten, DeepInfra, Fireworks, FriendliAI, OpenRouter, and Together AI, among others.
For enterprise environments, NVIDIA adds its own solution: NVIDIA NIM, a microservice for deployment “on any infrastructure accelerated by NVIDIA,” offering the same focus on privacy and control.
At the same time, they highlight early deployments in companies and consultancies (from integrators to software platforms), indicating that the model is designed for real-world workflows, not just demos.
What This Means for the Market (and Why Now)
Nemotron 3 arrives at a time when many organizations have uncovered a hard truth: agents work, but at a cost. Every layer of reasoning, tool integration, and verification adds tokens, latency, and operational risk. If the goal is to move from pilots to production, efficiency is no longer a technical detail but a budgetary requirement.
NVIDIA believes that the intermediate path—efficient open models fine-tuned with datasets and RL, coexisting with proprietary models when needed—will be the dominant pattern by 2026. Their clear strategy: to deliver an open stack with competitive performance, training tools, and a sovereignty/control narrative aligned with regulations and national strategies.
FAQs
What is Nemotron 3 Nano and what tasks is it suited for?
It’s the “small” model in the family (30 billion parameters, with partial MoE activation) aimed at efficient tasks like summarization, retrieval, assistant workflows, and supporting multi-agent systems with controlled costs.
What does a context window of 1 million tokens mean in an open model?
It allows maintaining long, multi-step flows (e.g., agents working with extensive documentation or large histories) with less need to fragment context or summarize aggressively—which can often reduce accuracy.
When will Nemotron 3 Super and Ultra be available?
NVIDIA targets the first half of 2026 for Super and Ultra.
What tools has NVIDIA released for training and evaluating agents using Nemotron 3?
Alongside datasets (including a security set for agents), the company launches NeMo Gym, NeMo RL, and NeMo Evaluator, available on GitHub and Hugging Face, with ecosystem tool support.