NVIDIA publishes Nemotron 3 Ultra, a 550B-parameter sparse MoE model

NVIDIA Research published Nemotron 3 Ultra on June 4, 2026, describing it as a 550B-total-parameter, 55B-active-parameter sparse mixture-of-experts model for advanced language-model workloads.

According to the NVIDIA Research project page, Nemotron 3 Ultra uses a hybrid Mamba-attention architecture. NVIDIA’s accompanying materials position the model around reasoning, long-context analysis, tool use, and complex agentic workflows, while reporting benchmark and throughput comparisons against GLM-5.1, Kimi-K2.6, and Qwen-3.5.

What NVIDIA says it released

The model is listed on Hugging Face as NVIDIA-Nemotron-3-Ultra-550B-A55B-NVFP4. The model card identifies it as a 550B-total, 55B-active model and gives the same June 4, 2026 release date. The “A55B” naming indicates that only a subset of the total parameters is active for a given inference pass, consistent with sparse mixture-of-experts design.

The Hugging Face listing also describes the release as an NVFP4 model. NVIDIA’s materials do not, in the provided excerpts, fully define the deployment implications of that format, but the model name and card make clear that NVIDIA is emphasizing a low-precision variant intended for efficient inference.

Architecture and evaluation details

NVIDIA’s technical report is the primary source for the model’s architecture, training, evaluation, and benchmark details. The report, as described by NVIDIA Research, provides the detailed basis for the company’s claims about the model’s design and performance.

The headline architectural point is the combination of mixture-of-experts sparsity with a hybrid Mamba-attention backbone. In practical terms, NVIDIA is presenting Nemotron 3 Ultra as a very large model by total parameter count, while keeping per-token active parameters far lower than the full 550B total.

NVIDIA Research also reports benchmark and throughput comparisons with GLM-5.1, Kimi-K2.6, and Qwen-3.5. Those comparisons should be read as NVIDIA-reported results unless independently reproduced, since the provided sources are NVIDIA’s own project page, model card, and technical report.

Why it matters

Nemotron 3 Ultra continues NVIDIA’s push to pair model releases with hardware- and inference-oriented claims. The model’s sparse 55B-active design is notable because it targets a familiar tradeoff in large language models: increasing total capacity while limiting the amount of computation used for each token.

The release also reflects a broader pattern in current frontier-model development, where companies are publishing models optimized not only for chat or general text generation, but also for tool use, long-context work, and multi-step reasoning tasks. NVIDIA’s Hugging Face model card explicitly names those intended use cases for Nemotron 3 Ultra.

For developers and researchers, the most important next step is likely independent evaluation. NVIDIA has provided the model listing and a technical report, but external testing will be needed to assess how its reported benchmark and throughput results translate across hardware setups, inference frameworks, prompting styles, and real-world workloads.