NVIDIA Nemotron 3 Ultra Explained: The 550B Open-Weight Model Built for AI Agents

NVIDIA's biggest open model is not new this week — and that is exactly why it is worth a clear-eyed explainer. Nemotron 3 Ultra was teased by CEO Jensen Huang on the Computex keynote stage in Taipei in early June 2026, and its weights landed on Hugging Face on June 4, 2026. It is the largest and final member of the Nemotron 3 family: a 550-billion-parameter mixture-of-experts model designed for long-running AI agents, released with its weights, training data and recipes in public. Social media clips are still circulating claims about it — a million-token context window, five times faster than rivals, an open-source answer to Claude and ChatGPT. Some of that is true. Some of it is NVIDIA's own marketing arithmetic. Here is what the primary sources actually say.

What NVIDIA actually released

The Nemotron 3 family rolled out in stages: Nano (31.6B parameters) shipped on December 15, 2025, Super (about 120B) followed at GTC on March 11, 2026, and Ultra closed the family out on June 4, 2026. All three share the same design philosophy — hybrid Mamba-attention architecture, mixture-of-experts routing, and context windows of up to 1 million tokens.

Ultra itself ships in four checkpoints on Hugging Face: the post-trained NVFP4 build (the flagship, quantized for NVIDIA's Blackwell hardware), a full-precision BF16 build, the base pre-trained model, and the GenRM model used in its reinforcement-learning pipeline. Alongside the weights, NVIDIA published training datasets — including a code corpus of 173 billion tokens drawn from GitHub through September 30, 2025 — plus post-training datasets and the full training recipe in its Nemotron developer repository. According to NVIDIA's launch announcement, the model was post-trained for agent platforms including Hermes Agent, LangChain Deep Agents, OpenClaw, OpenHands and OpenCode.

The headline specifications, from NVIDIA's official model card:

Two design choices explain most of the marketing claims. First, sparsity: only about a tenth of the parameters fire on any token, so inference cost tracks a ~55B model while knowledge capacity sits at 550B scale. Second, NVFP4: the model was pre-trained in NVIDIA's 4-bit floating-point format for Blackwell GPUs, and its MTP layers let it draft several tokens at once (native speculative decoding). It also supports inference-time reasoning budget control — operators can dial how much the model "thinks" up or down, and reasoning can be switched on or off via the chat template.

What the benchmarks actually say

Every number in the table below is NVIDIA's own, published on the model card and collected with NVIDIA's NeMo Evaluator tooling. Treat them as the manufacturer's figures, not independent results:

On speed, NVIDIA's research page claims 5.9x, 4.8x and 1.6x higher inference throughput than GLM-5.1 (754B), Kimi K2.6 (1T) and Qwen 3.5 (397B) respectively — in one specific configuration, 8K input tokens and 64K output tokens. NVIDIA's launch materials also cite more than 300 output tokens per second and roughly 30% lower cost on complex agentic tasks. Those are the source of the "5x faster" line in the viral clips: real numbers, but vendor numbers, measured on NVIDIA's preferred hardware and workload shape.

The independent check comes from Artificial Analysis, whose Intelligence Index scores Ultra at 47.7 (NVFP4) and 48.2 (BF16). That is the highest score the tracker gives any US-built open-weight model — ahead of Google's Gemma 4 31B (39.2) and NVIDIA's own Super (36.0). It is not the top of the open table: Moonshot's Kimi K2.6 (53.9), DeepSeek V4 Pro (52.0) and Zhipu's GLM-5.2 (51.0) all score higher, and closed frontier models such as Claude Opus 4.8 (61.4) and GPT-5.5 (60.2) sit in a different league. NVIDIA has also reported, in a September 2026 Hugging Face post, that systems built on Ultra checkpoints reached gold-medal-level results at the IOI 2026 programming olympiad (535.4 of 600 points) and the IMO 2026 maths olympiad (30 of 42) — specialized competition systems built on the base model, not the stock download, but a genuine signal of the base model's reasoning ceiling.

The hardware reality

This is the part the short clips skip. Nemotron 3 Ultra is open-weight, not lightweight. NVIDIA's model card lists a minimum of four B200-class data-center GPUs for a single node of the NVFP4 build, and its serving examples use vLLM across multi-GPU nodes. There is no version of this model that runs on a laptop, a gaming RTX card, or a single workstation GPU.

That does not make the release irrelevant to smaller teams — it changes how you would use it:

• Hosted endpoints: try it through NVIDIA's build.nvidia.com catalog or a third-party inference provider rather than buying hardware; launch coverage reported some providers serving it at 300+ tokens per second.
• Self-hosting for enterprises: the real audience is organizations with Blackwell clusters that want frontier-adjacent capability on their own infrastructure for data-residency or compliance reasons — the same pitch as Meta's Muse Spark Open, at somewhat smaller scale.
• Local Nemotron, smaller size: if the goal is an agent on your own machine, NVIDIA's intended answers are Nemotron 3 Nano and the Nemotron 3.5 Super VL and Bolt models, which go down to a 4GB footprint for edge devices.

Honest limitations

• The speed crown is self-awarded. The throughput multipliers and the 300-tokens-per-second figure are NVIDIA launch claims. Independent serving results vary by provider and workload, and Artificial Analysis's composite still ranks three Chinese open models as more intelligent.
• Text only. Unlike the separate Nemotron 3.5 Super VL, Ultra takes text in and returns text out — no image or document understanding in this checkpoint.
• "Open" has a license attached. OpenMDW-1.1 permits commercial use per the model card, but it is NVIDIA's own license, not Apache 2.0. Product teams need to read it, the same caution that applies to Mistral Large 4 and other vendor-licensed open-weight releases.
• Benchmarks are a snapshot. NVIDIA's scores were collected in its own harness, and third-party leaderboards keep moving — one independent index (Vals.ai) placed Ultra fifth overall in its August 2026 update, a reminder that launch-day rankings decay fast.

What this signals

Nemotron 3 Ultra is best understood as NVIDIA selling the picks and shovels by giving away the map. A 550B open model that runs best in NVFP4 on Blackwell GPUs is a demand engine for exactly the hardware NVIDIA sells — and by publishing the data and recipes too, NVIDIA seeds a generation of fine-tunes that all assume its tooling. The strategic context is the open-weight race: the US now has a credible, downloadable flagship of its own, even if Kimi K2.6 still tops the independent index and the closed frontier remains out of reach. For everyone else, the practical takeaway is narrower and more useful: the strongest open models are now genuinely agent-grade at coding, tool use and long-context work — Kimi's open-weight line included — but "free to download" and "cheap to run" remain very different sentences.

Sources & Further Reading

• Nemotron 3 Ultra — NVIDIA Research (official model page)
• NVIDIA-Nemotron-3-Ultra-550B-A55B-NVFP4 — official model card, Hugging Face
• Nemotron 3 family — NVIDIA Research
• NVIDIA Nemotron 3 Ultra: 550B Open-Weight AI — Tech Insider analysis
• NVIDIA Nemotron 3 hits gold at IMO and IOI 2026 — coverage of NVIDIA's Hugging Face report

All benchmark figures are NVIDIA's own unless attributed to Artificial Analysis or another independent source. The featured image is an original illustration created for this article, not a photograph.