Nvidia announced Nemotron 3 Ultra on June 4, 2026, positioning the open-source flagship model for long-running AI agents and complex multi-agent workflows. The model carries 550 billion total parameters and uses a Mixture-of-Experts architecture, while activating only 55 billion parameters per inference step. The stated goal is direct: cut the cost and efficiency drag created by long context windows, repeated planning, tool calls, and sub-agent coordination.
Built as an orchestrator for multi-agent workloads
Nvidia describes Nemotron 3 Ultra as an orchestrator or frontier reasoning engine inside agent stacks. In that setup, the model handles deep planning, complex synthesis, and logical verification, while lighter models take on routine execution and tool usage. The source material argues that conventional multi-agent workflows burn through tokens as they keep revisiting plans, managing long context, and delegating tasks, pushing both cost and task drift higher.
Benchmark results were presented around agent-focused workloads. Nemotron 3 Ultra scored 91% on PinchBench, while posting 40% on EnterpriseOps-Gym and 67% on Terminal-Bench 2.0. Nvidia says the model’s overall reasoning performance surpasses or matches major open models such as GLM 5.1, Kimi K2.6, and Qwen3.5, even though only a fraction of its total parameters are active during computation.
Five technical changes aimed at speed and lower cost
The company tied the model’s performance gains to five core components: hybrid Mamba-Transformer layers, NVFP4 quantization, LatentMoE, multi-token prediction (MTP), and multi-teacher online distillation (MOPD). The hybrid Mamba-Transformer design is meant to combine efficient long-sequence processing with accurate factual recall. NVFP4 quantization allows the same model weights to run across Hopper, Blackwell, and Ampere GPU architectures.
According to the source, Nemotron 3 Ultra can deliver up to 5x higher throughput on Blackwell compared with the traditional BF16 format. Nvidia also says the model cuts total token consumption during task execution, which can reduce enterprise agent-task costs by as much as 30%. The pitch is simple: faster output, lower operating cost.
Open weights, training recipe, and deployment access
On training data, Nemotron 3 Ultra is built on more than 10 trillion tokens of pretraining and adds over 212 billion domain-specific tokens, including legal documents, Wikipedia-style text, and recent GitHub code. Nvidia said the model is fully open source under the OpenMDW-1.1 license, with model weights, training recipe, and data pipelines made available to the community.
Developers can already access and deploy Nemotron 3 Ultra through Hugging Face, NVIDIA Build, and NIM. The source also cited a 95% result on Ruler @1M, highlighting long-context handling as a core strength. Nvidia pointed to enterprise uses such as customer service automation, supply chain management, IT security, and chip design verification.