Nemotron 3 Ultra

Overview

Nemotron 3 Ultra is the flagship tier of NVIDIA's Nemotron 3 family of open models, released on June 4, 2026. It is a 550-billion-parameter Mixture-of-Experts large language model that activates 55 billion parameters per token — roughly 10% sparsity — built on a hybrid Mamba-2 / Transformer architecture with a 1M-token context window.

Nemotron 3 is a family, not a single model. NVIDIA debuted it on December 15, 2025 with the small Nano tier, added the mid-size Super at GTC on March 11, 2026, and closed the line with Ultra at Computex in June. All three share the same architecture and the same million-token context, and differ only in scale:

• Nano — 31.6B total / 3.2B active — cost-efficient sub-agents, retrieval, and edge deployment
• Super — 120B total / 12B active — high-throughput multi-agent production workloads
• Ultra — 550B total / 55B active — frontier reasoning, deep research, and planning

Licensing is worth reading closely, because it is not uniform across the family. Ultra ships under the OpenMDW License Agreement v1.1 — the Linux Foundation's permissive model license, which covers weights, code, documentation, and training data as one redistributable whole. The smaller Nano and Super tiers use the NVIDIA Nemotron Open Model License instead. Either way NVIDIA's stated posture is unusually open: it says it is releasing "the model weights, training recipe, and all the data for which we hold redistribution rights," so the recipe and data travel with the models, not only the parameters.

The second theme is efficiency. NVIDIA positions Ultra less as a benchmark-topper and more as the most throughput-efficient frontier open model, achieved through three architectural bets — a hybrid Mamba-Transformer backbone, LatentMoE routing, and Multi-Token Prediction — and NVFP4 pretraining that targets its own Blackwell hardware.

Capabilities

• Frontier open-weight reasoning: Ultra is NVIDIA's highest-accuracy open reasoning model, aimed at deep research, multi-step planning, and long-running agent loops.
• Million-token context: A 1M-token window across the whole family, with NVIDIA reporting strong RULER accuracy for Ultra at the full 1M length rather than only at short contexts.
• Hybrid Mamba-Transformer throughput: Interleaving linear-time Mamba-2 layers with a minority of attention layers gives better tokens-per-GPU economics than a pure Transformer of similar accuracy.
• LatentMoE routing: Experts are routed in a compressed latent space rather than the full hidden dimension, which NVIDIA describes as fitting more experts into the same inference budget than a conventional MoE.
• Native speculative decoding: Multi-Token Prediction (MTP) layers are trained in, giving faster generation without a separate draft model.
• NVFP4 from pretraining: Super and Ultra are pre-trained in NVFP4 4-bit floating point, so the quantized build is a first-class artifact, not a lossy afterthought.
• Self-hostable and inspectable: openly licensed weights, recipe, and data (Ultra under OpenMDW-1.1) allow private deployment, fine-tuning, and auditing.

Technical Specifications

• Architecture: hybrid Mamba-2 / Transformer Mixture-of-Experts with LatentMoE routing and Multi-Token Prediction layers
• Pretraining precision: NVFP4 (Super and Ultra); Blackwell-native 4-bit floating point
• Ultra model ID: nvidia/nemotron-3-ultra-550b-a55b on hosted APIs
• Weights: published on Hugging Face — see the checkpoints below
• License: Ultra under OpenMDW License Agreement v1.1 (weights, recipe, and redistributable data); Nano and Super under the NVIDIA Nemotron Open Model License

Ultra checkpoints

NVIDIA ships Ultra as several checkpoints rather than one file:

• NVIDIA-Nemotron-3-Ultra-550B-A55B-BF16 — post-trained instruct model, ~561B params including embeddings
• NVIDIA-Nemotron-3-Ultra-550B-A55B-Base-BF16 — the pre-trained base, for teams doing their own post-training
• NVIDIA-Nemotron-3-Ultra-550B-A55B-NVFP4 — the 4-bit build (~335B), the practical self-hosting target on Blackwell
• NVIDIA-Nemotron-3-Ultra-550B-A55B-GenRM — a generative reward model fine-tuned from the BF16 base, for scoring and ranking responses during RLHF, not a chat model

Nano Omni and the small Nano

Two Nano variants sit outside the three headline tiers. Nano Omni (Nemotron-3-Nano-Omni-30B-A3B-Reasoning) is the family's multimodal member, accepting more than text and shipping in BF16, FP8, and NVFP4. A separate small dense Nano 4B targets tight edge and on-device budgets. The core Nano 30B-A3B, Super, and Ultra tiers are text-only.

Use Cases

• Deep research and strategic planning: Ultra's positioning — long context plus frontier reasoning — suits multi-document synthesis and analyst-style workflows that read broadly before concluding.
• Long-running autonomous agents: MTP-driven speculative decoding and hybrid-MoE throughput keep multi-step tool-using loops affordable where a dense frontier model would be too slow.
• On-premises and regulated deployment: OpenMDW weights make air-gapped, private hosting viable for finance, healthcare, and public-sector work that cannot send data to a third-party API.
• Fine-tuning and distillation: Because the base checkpoints, recipe, and data are published, Super and Ultra are practical bases for domain adaptation, and the GenRM checkpoint slots directly into an RLHF pipeline.
• Cost-tiered fleets: Route bulk sub-agent traffic to Nano, coordinated production agents to Super, and only the hardest reasoning to Ultra — one architecture across the tier reduces integration cost.
• 1M-context retrieval and analysis: Whole-repository review, retrieval-augmented pipelines over large corpora, and long-transcript analysis that would otherwise need aggressive chunking.

Performance / Benchmarks

NVIDIA frames Ultra on efficiency and parity rather than a benchmark sweep. Its research page states Ultra "achieves on-par accuracies compared to other state-of-the-art open LLMs across a diverse set of benchmarks," and singles out long-context accuracy on RULER at 1M tokens as a strength.

The headline claim is throughput. NVIDIA reports Ultra as substantially faster per GPU than the open frontier models it compares against — naming GLM-5.1-754B-A40B, Kimi-K2.6-1T-A32B, and Qwen-3.5-397B-17B in its throughput charts — attributing the gap to the hybrid Mamba-Transformer backbone, LatentMoE, and NVFP4 execution on Blackwell. NVIDIA makes the same efficiency case for Super, which it positions as the throughput leader in the ~120B open class.

Full per-benchmark tables (AIME, GPQA, LiveCodeBench, and the rest) are in the Nemotron 3 Ultra Technical Report, dated June 9, 2026. As with any first-party release, these are vendor-selected numbers; treat them as an upper bound and validate on your own evaluation set — independent third-party replication of the Nemotron 3 scores is still limited.

Limitations

• No single "Nemotron 3" model: You must pick a tier and a checkpoint. Ultra's callable ID is nvidia/nemotron-3-ultra-550b-a55b; there is no generic nemotron-3 endpoint.
• Serving cost at 550B: Even at 55B active parameters, self-hosting Ultra requires a large multi-GPU cluster. The NVFP4 build (~335B) lowers the bar but still targets Blackwell-class hardware; Super or Nano are the realistic self-host targets for most teams.
• Output caps on hosted endpoints: The 1M figure is the input context. Hosted providers cap maximum output far lower (OpenRouter's paid Ultra endpoint tops out around 16K output tokens), which constrains single-response long-form generation.
• Text-only core: Nano, Super, and Ultra are text models. Multimodal input requires the separate Nano Omni, which is only offered at the 30B-A3B size — there is no Ultra-scale multimodal tier.
• No fixed vendor pricing: Being open-weight, per-token cost depends entirely on which host you use; prices and available context/output limits vary by provider.
• Vendor-reported benchmarks: NVIDIA's accuracy and throughput claims are self-published and, for Ultra, recent — independent verification is thin.

Pricing & Access

Nemotron 3 Ultra is open-weight, so cost depends on how you run it. The weights are free to download; hosted access is priced by each provider.

Access options:

• Hugging Face — BF16, NVFP4, base, and GenRM checkpoints under OpenMDW-1.1
• NVIDIA build / NIM — hosted preview and enterprise NIM microservices
• OpenRouter and other aggregators — OpenAI-compatible hosted API, free and paid tiers

Ecosystem & Tools

• NVIDIA Nemotron 3 research page — family overview, architecture, and checkpoints
• Nemotron 3 Ultra research page — the flagship tier's details and comparisons
• NVIDIA Nemotron v3 collection on Hugging Face — every Nano, Super, Ultra, and Omni checkpoint
• NVIDIA Nemotron developer hub — NIM deployment, docs, and datasets
• NeMo framework — NVIDIA's training and fine-tuning stack, aligned with the released recipe and data

Community & Resources

• NVIDIA Newsroom — Nemotron 3 family debut (December 15, 2025)
• Nemotron 3 Ultra Technical Report (PDF) (June 9, 2026)
• Nemotron 3 Ultra on OpenRouter — hosted API, pricing, and providers
• Compare with DeepSeek V4, GLM-5.2, Kimi K3, and Qwen3.7-Max