Availability: While the Nano model is available immediately, larger Super and Ultra variants are not scheduled until the first half of 2026.
NVIDIA is breaking from standard AI architecture with the launch of its Nemotron 3 family, introducing a hybrid design that combines Mamba-2 state-space models with traditional Transformers to slash inference costs.
Targeting the bottleneck in "agentic" AI workflows, the new architecture uses a Mixture-of-Experts (MoE) approach to activate only a fraction of its parameters per token. This design allows the entry-level Nemotron 3 Nano to deliver four times the throughput of its predecessor while reducing memory usage.
Available immediately, the 30-billion-parameter Nano model is the first of three planned releases. Larger "Super" and "Ultra" variants, designed for complex reasoning tasks, are scheduled to arrive in the first half of 2026.
Breaking from pure Transformer designs, Nemotron 3 integrates Mamba-2 state-space models (SSMs) with traditional attention layers. Designed to solve the quadratic complexity of Transformers, where memory usage scales poorly with long sequences, the architecture offers a distinct efficiency advantage.
By interleaving Mamba layers, the model reduces the need to store extensive key-value caches, directly lowering VRAM requirements. Targeting the "KV cache" bottleneck, a major cost driver in long-context inference, the design enables higher throughput on existing hardware.
Kari Briski, Vice President for Generative AI Software at NVIDIA, noted that "the hybrid Mamba transformer architecture runs several times faster with less memory, because it avoids these huge attention maps and key value caches for every single token."
A Mixture-of-Experts (MoE) design further optimizes efficiency, activating only a subset of parameters for each token. For the Nano model, this results in just 3.5 billion active parameters out of a total 30 billion, keeping latency low.
Performance metrics show a 4x increase in token throughput compared to the previous generation Nemotron 2 Nano. Reasoning-token generation costs are slashed by 60%, a critical metric for agentic workflows that require extensive "thinking" steps.
Early benchmarks from Artificial Analysis benchmarks rank the Nano model as the "most open and efficient" in its specific size class. However, the delay for the larger models leaves a gap that competitors like Meta and Mistral may continue to exploit.
NVIDIA is positioning the Nemotron 3 family not as general-purpose chatbots, but as specialized components in a multi-agent system. Reflecting a "router-reasoner" paradigm, the lineup features three distinct sizes: Nano, Super, and Ultra.
Nano (30B) serves as the high-speed edge node or router, handling simple tasks and dispatching complex queries. Perplexity's adoption highlights this routing strategy, using open models to optimize costs before calling expensive proprietary frontiers.
A substantial 1 million token context window allows the Nano model to ingest entire codebases or document libraries in a single pass. Such capacity is essential for Retrieval-Augmented Generation (RAG), where agents must synthesize answers from vast external data.
The "Super" (100B) and "Ultra" (500B) models are designed to handle the heavy lifting of complex reasoning once routed. By specializing models, NVIDIA aims to reduce the "communication overhead" and latency often seen in multi-agent orchestrations.
Beyond the models themselves, NVIDIA is releasing the tooling required to train and fine-tune them, specifically NeMo Gym. NeMo Gym provides a reinforcement learning (RL) environment, allowing developers to train agents through trial and error rather than just supervised fine-tuning.
Open-sourcing the training environment represents a strategic play to keep developers within the NVIDIA software stack. Jensen Huang, Founder and CEO of NVIDIA, stated that "with Nemotron, we're transforming advanced AI into an open platform that gives developers the transparency and efficiency they need to build agentic systems at scale."
Supporting "Sovereign AI," the initiative enables nations and enterprises to build models on their own infrastructure using local data. This localization is critical for regulated industries (finance, healthcare) and regions with strict data sovereignty laws like the EU.
By offering these tools, NVIDIA ensures that even "open" model development drives demand for its Blackwell and H100 hardware. Strategically, this creates a sticky ecosystem where the software (NIM, NeMo) is free, but the compute required to run it effectively is proprietary.
Immediate availability is limited to the Nemotron 3 Nano model, which can be downloaded now via the Hugging Face repository. The larger reasoning engines, Super and Ultra, are not scheduled for release until the first half of 2026.
Staggering the release puts NVIDIA in a unique position: leading on architectural novelty but trailing on immediate high-parameter availability. In the interim, developers must rely on the Nano model for routing while using competitors (like Llama 3.1 405B) for heavy reasoning.
Addressing the challenges developers face with current open models, Briski observed that "most open models force developers into painful trade-offs between efficiencies like token costs, latency, and throughput."