NVIDIA Research and the NVIDIA Developer Blog introduced Nemotron 3 Ultra, a 550-billion-parameter total, 55-billion-active mixture-of-experts model designed for long-running agent workflows, with a reported context window of up to 1 million tokens and benchmark claims for throughput and long-context retrieval.
NVIDIA Research has introduced Nemotron 3 Ultra, an open 550-billion-parameter total, 55-billion-active model aimed at long-context reasoning and long-running agent workloads.
According to the NVIDIA Research project page for NVIDIA Nemotron 3 Ultra, the model is positioned as an open model with support for context windows of up to 1 million tokens. NVIDIA says the model is designed for agentic reasoning workloads where systems may need to retain and use large amounts of information across extended sessions.
The accompanying NVIDIA Research technical report, titled “Nemotron 3 Ultra: Open, Efficient Mixture-of-Experts Hybrid Mamba-Transformer Model for Agentic Reasoning,” describes the model as a mixture-of-experts hybrid Mamba-Attention architecture with 550 billion total parameters and 55 billion active parameters. The report says the model was trained on 20 trillion tokens and then post-trained using supervised fine-tuning, reinforcement learning, and MOPD, a post-training method referenced in the report.
NVIDIA’s Developer Blog frames the release around “long-running agents,” emphasizing reasoning, throughput, and efficiency rather than only raw model scale. The blog says Nemotron 3 Ultra incorporates techniques including NVFP4, LatentMoE, and multi-token prediction.
NVIDIA Research says Nemotron 3 Ultra supports up to a 1 million-token context window and reports RULER benchmark results at that context length. RULER is commonly used to test whether long-context models can retrieve and use information placed deep inside very large input sequences, though benchmark results do not necessarily capture all production use cases.
The NVIDIA Research page also says Nemotron 3 Ultra achieved higher throughput than GLM-5.1, Kimi-K2.6, and Qwen-3.5 in NVIDIA’s tested setting. That comparison should be read with the usual caveat that throughput depends heavily on hardware, serving configuration, sequence length, batch size, precision, and implementation details. NVIDIA’s claim is therefore best understood as a reported result within its own evaluation setup, not a universal ranking across all deployments.
The technical report’s description of Nemotron 3 Ultra as a hybrid Mamba-Transformer model is notable because it reflects a broader trend in large-model research: combining attention-based components with alternative sequence-modeling mechanisms to improve efficiency at long context lengths. NVIDIA’s use of a mixture-of-experts design also means that only a subset of the model’s parameters are active for a given token, which can reduce compute requirements compared with activating the full parameter count on every forward pass.
NVIDIA’s stated 55 billion active parameters figure is central to that efficiency argument. It means the model’s total capacity is much larger than the number of parameters used per token, although real-world cost and speed still depend on system design and deployment hardware.
NVIDIA describes Nemotron 3 Ultra as an open model, and its announcement materials focus on enterprise and developer use cases involving extended reasoning, tool use, and persistent agent behavior. The Developer Blog specifically says the model is intended to power faster and more efficient reasoning for long-running agents.
The release adds to NVIDIA’s expanding Nemotron model family and reinforces the company’s effort to pair model releases with its software and hardware stack. The most important published details so far are the model’s reported scale, its hybrid MoE architecture, its long-context support, and NVIDIA’s own benchmark claims for RULER performance and throughput.
As with any vendor-published model announcement, independent evaluations will be important for assessing how Nemotron 3 Ultra performs outside NVIDIA’s chosen benchmarks and serving conditions.
NVIDIA Research has introduced Nemotron 3 Ultra, an open 550 billion parameter total, 55 billion active model aimed at long context reasoning and long running agent workloads.
What NVIDIA announced According to the NVIDIA Research project page for NVIDIA Nemotron 3 Ultra , the model is positioned as an open model with support for context windows of up to 1 million tokens .
NVIDIA says the model is designed for agentic reasoning workloads where systems may need to retain and use large amounts of information across extended sessions.
Continue reading
