Nemotron 3 Ultra

NVIDIA's open Nemotron line: the 550B/55B-active Nemotron 3 Ultra flagship and Nemotron 3.5 Lightning, a 30B/3B-active agent-execution model, both OpenMDW.

Last reviewed September 6, 2026

At a glance

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 is a line, not a single model. NVIDIA debuted Nemotron 3 on December 15, 2025 with the small Nano tier, added the mid-size Super at GTC on March 11, 2026, and topped it out with Ultra at Computex in June. On August 11, 2026 it added Nemotron 3.5 Lightning, a 30B / 3B-active successor to Nano aimed squarely at the execution layer of agent workloads. All four share the same hybrid architecture and the same million-token ceiling, and differ in scale and in what they are for:

• Nemotron 3 Nano — 31.6B total / 3.2B active — cost-efficient sub-agents, retrieval, and edge deployment
• Nemotron 3 Super — 120B total / 12B active — high-throughput multi-agent production workloads
• Nemotron 3 Ultra — 550B total / 55B active — frontier reasoning, deep research, and planning
• Nemotron 3.5 Lightning — 30B total / ~3B active — high-volume, low-latency agent execution steps

The “3.5” label is a generation bump on one model, not a new size ladder: as of September 2026 there is no Nemotron 3.5 Super or Ultra, and Ultra remains the top of the line. (A separately-scoped Nemotron 3.5 Content Safety guard model — 4B, fine-tuned from Gemma 3 4B — shares the version label but neither the architecture nor the license.)

Licensing is worth reading closely, because it is not uniform. Ultra and Lightning ship 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 older 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. Lightning applies the same bets one tier down, where they matter most: an agent loop that makes a thousand tool calls pays for latency a thousand times.

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.
• Fast agent execution: Lightning is built for the steps between the reasoning — tool calls, result validation, extraction, sub-agent delegation — where NVIDIA claims up to 4x the output speed of similar-sized open models.
• Million-token context: A 1M-token window across the whole line, 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; Lightning interleaves Mamba-2, MoE and select attention layers the same way.
• LatentMoE routing: Experts are routed in a compressed latent space rather than the full hidden dimension, which NVIDIA says fits more experts into the same inference budget than a conventional MoE.
• Native speculative decoding: Multi-Token Prediction layers are trained in, so generation is faster without a separate draft model; Lightning adds purpose-built DSpark and DFlash draft checkpoints.
• NVFP4 from pretraining: Super, Ultra and Lightning 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 and Lightning under OpenMDW-1.1) allow private deployment, fine-tuning, and auditing — and 3B active parameters at 4-bit put a 1M-context model on a single device.

Technical Specifications

• Architecture: hybrid Mamba-2 / Transformer Mixture-of-Experts with LatentMoE routing and Multi-Token Prediction layers
• Pretraining precision: NVFP4 (Super, Ultra and Lightning); Blackwell-native 4-bit floating point
• Ultra model ID: nvidia/nemotron-3-ultra-550b-a55b on hosted APIs
• Lightning model ID: nvidia/nemotron-3.5-lightning on OpenRouter; NVIDIA-Nemotron-3.5-Lightning-30B-A3B on Hugging Face
• License: Ultra and Lightning 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:

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

Nemotron 3.5 Lightning

Released August 11, 2026, Lightning is a 30B hybrid MoE with roughly 3B active parameters, interleaving Mamba-2 and MoE layers with select attention layers. NVIDIA reports pre-training on more than 20 trillion tokens with an NVFP4 recipe, a pre-training data cutoff of September 2025 and a post-training cutoff of May 2026. It is licensed OpenMDW-1.1, commercial use permitted.

Checkpoints:

• NVIDIA-Nemotron-3.5-Lightning-30B-A3B-BF16 — reference full-precision weights, the customization starting point
• NVIDIA-Nemotron-3.5-Lightning-30B-A3B-NVFP4 — the production inference build
• …-NVFP4-DSpark and …-NVFP4-DFlash — speculative-decoding checkpoints
• Official GGUF weights for local runtimes (Ollama, LM Studio, llama.cpp)

Hardware. NVIDIA validates single-device deployment on DGX Spark (GB10) at the full 1M context, on GB200 with native FP4 tensor cores, and on 1x H100 80GB or 1x A100 80GB (the A100 via W4A16); H100/H200 scale to 8 GPUs. The BF16 card notes that a single H100 80GB is memory-bound to roughly 256K tokens of context. For local desktop use the RTX 5090 is a named target, and Unsloth’s GGUF conversion puts Q4_K_M at 25.3GB — comfortable on a 32GB card, above the ceiling of every 24GB one.

Nano Omni and the small Nano

Two Nano variants sit outside the headline tiers. Nano Omni (Nemotron-3-Nano-Omni-30B-A3B-Reasoning) is the line’s multimodal member, shipping in BF16, FP8, and NVFP4; a separate dense Nano 4B targets tight edge budgets. Nano 30B-A3B, Super, Ultra and Lightning are text-only.

Use Cases

The first decision on this page is which model to route to, and the answer is not “the biggest one you can afford”. Ultra and Lightning are built for different halves of the same agentic workflow.

Route to Nemotron 3 Ultra when the step needs to think:

• Deep research and planning: long context plus frontier reasoning suits multi-document synthesis and analyst work that reads broadly before concluding.
• Orchestration and plan repair: deciding what the agent does next, and recovering when a plan fails — the steps where a wrong answer costs more than the tokens.
• Fine-tuning and distillation: base checkpoints, recipe and data are published, and GenRM slots into an RLHF pipeline.

Route to Nemotron 3.5 Lightning when the step is high-volume and specified:

• Always-on agent execution: tool calls, argument construction, result validation, and sub-agent delegation, where a frontier model adds latency and cost to a decision that was never hard.
• Classification, extraction and routing: the per-item work in a pipeline, at roughly a quarter of Ultra’s hosted input price.
• Local and on-device agents: an OpenMDW model with a 1M-token ceiling that runs on one workstation GPU or a DGX Spark, for work that cannot leave the machine.

Applies to both:

• On-premises and regulated deployment: OpenMDW weights make air-gapped, private hosting viable for work that cannot reach a third-party API.
• 1M-context retrieval and analysis: whole-repository review, retrieval-augmented pipelines, and long-transcript work that would otherwise need aggressive chunking — subject to the hosted context caps below.
• Cost-tiered fleets: bulk execution to Lightning, general production traffic to Super, the hardest reasoning to Ultra. NVIDIA’s NeMo Switchyard router automates exactly this split.

Performance / Benchmarks

Ultra. 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 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 in its charts — GLM-5.1-754B-A40B, Kimi-K2.6-1T-A32B, Qwen-3.5-397B-17B — crediting the hybrid backbone, LatentMoE, and NVFP4 execution on Blackwell. Per-benchmark tables are in the Nemotron 3 Ultra Technical Report, dated June 9, 2026.

Lightning. These are the scores NVIDIA published on the Lightning model cards at release (August 11, 2026) — vendor-run, self-reported, on both the reference and production checkpoints:

NVIDIA’s speed claims are the point of the model rather than the accuracy ones: up to 4x the output speed of similar-sized models, and 10,000 PinchBench tasks completed 30% faster than Qwen3.6 35B at comparable accuracy. Note the gap between those two numbers — a 4x token-generation speedup buying a 30% task speedup is a fair illustration that orchestration, not generation, is the bottleneck in agent loops.

Third-party measurement is available for Lightning, which is more than Ultra can say. Artificial Analysis measures a median 284.6 output tokens/s and 1.01s time to first token across providers, at a median $0.07 input / $0.22 output per 1M tokens. It scores the model 16 on Intelligence Index v4.2; its launch-day write-up put Lightning at 24, on the index version current at the time of publication, ahead of Nemotron 3 Nano and level with gpt-oss-120b. Those two figures are not comparable with each other, which is precisely why the index version has to travel with the score.

As with any first-party release, NVIDIA’s tables are vendor-selected numbers; treat them as an upper bound and validate on your own evaluation set.

Limitations

• Serving cost at 550B: Even at 55B active, self-hosting Ultra needs a large multi-GPU cluster. The NVFP4 build (~335B) lowers the bar but still targets Blackwell; Super, Nano or Lightning are the realistic self-host targets for most teams.
• “Runs on one GPU” is not “runs on a laptop”: NVIDIA’s named local targets for Lightning are RTX PCs, RTX PRO workstations, DGX Spark, DGX Station and Jetson — desktop and embedded hardware. The default Q4_K_M GGUF is 25.3GB, above every 24GB mobile card, and the full 1M context on a single H100 is only reachable in the NVFP4 build.
• Hosted context is a quarter of the headline: The 1M figure is the model’s native ceiling. OpenRouter serves both Ultra and Lightning at 262K, and caps Ultra’s output at 16,384 tokens.
• Lightning is not a reasoning model: NVIDIA positions it as a specialized execution layer. Its Terminal-Bench and Intelligence Index placements sit well short of frontier, and routing hard planning steps to it will cost more in retries than it saves in tokens.
• No Nemotron 3.5 size ladder: There is no 3.5 Super or Ultra. If you standardize on the 3.5 generation you are standardizing on one 30B model, with Nemotron 3 Ultra — a different generation — as your only in-family escalation path.
• Text-only core: Multimodal input requires the separate Nano Omni, offered only at 30B-A3B — there is no Ultra-scale multimodal tier.
• No fixed vendor pricing: per-token cost depends entirely on which host you use, and context and output limits vary with it.
• Vendor-reported benchmarks: NVIDIA’s accuracy and throughput claims are self-published, and independent replication of the Nemotron 3 scores remains thin.

Pricing & Access

Both models are open-weight, so cost depends on how you run them. The weights are free to download; hosted access is priced by each provider. Rates below were checked on September 6, 2026.

Access options:

• Hugging Face — BF16, NVFP4, GGUF, speculative-decoding, base, and GenRM checkpoints
• NVIDIA build / NIM — hosted preview and enterprise NIM microservices; ModelScope mirrors
• OpenRouter and other aggregators — OpenAI-compatible hosted API, free and paid tiers; Lightning is also served by Baseten, Together AI and Nebius
• Local runtimes — Ollama, LM Studio and llama.cpp for the Lightning GGUF builds

Ecosystem & Tools

• NeMo Switchyard — the open-source model router NVIDIA shipped alongside Lightning on August 11, 2026. It directs each step of an agent workflow to the cheapest model that can handle it, across open, proprietary and NVIDIA models, with tuning-free classifier, stage and escalation routers plus a learnable prefill router. NVIDIA’s reported LangChain benchmark escalates between Lightning and Claude Opus 4.8 at roughly a third of a frontier-only baseline’s cost. It answers “which Nemotron should I call” at runtime rather than at design time.
• The Nemotron Coalition — announced March 16, 2026: Black Forest Labs, Cursor, LangChain, Mistral AI, Perplexity, Reflection AI, Sarvam and Thinking Machines Lab, collaborating with NVIDIA on open frontier models. The first output is a base model co-developed by Mistral AI and NVIDIA, trained on DGX Cloud and open-sourced, which NVIDIA says will underpin the coming Nemotron 4 family — so the next generation is not purely NVIDIA-built.
• 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)
• NVIDIA Newsroom — Nemotron Coalition launch (March 16, 2026)
• Nemotron 3 Ultra Technical Report (PDF) (June 9, 2026)
• NVIDIA Technical Blog — Nemotron 3.5 Lightning (August 11, 2026)
• Nemotron 3.5 Lightning BF16 model card — benchmarks, hardware support, and license
• Nemotron 3 Ultra on OpenRouter — hosted API, pricing, and providers
• Compare with DeepSeek V4, GLM-5.3, Kimi K3, Gemma 4, and Qwen3.8-Max