Bottom line up front
Nemotron 3 Ultra is a 550-billion-parameter mixture-of-experts model with about 55 billion active per token, using a hybrid LatentMoE architecture of Mamba-2, MoE, and attention layers with multi-token prediction. NVIDIA documents up to a 1-million-token context, OpenMDW-1.1 licensing, NVFP4 pre-training, and inference-time reasoning budget control. The published minimum for self-hosting is eight H100-class GPUs, so the free hosted endpoint is the realistic access route for most teams — and it comes with the usual subsidy conditions.
A hybrid architecture built for long agent runs
NVIDIA describes Nemotron 3 Ultra as a LatentMoE model: a hybrid of Mamba-2, mixture-of-experts, and attention layers, with multi-token-prediction layers for native speculative decoding. About 55 billion of its 550 billion parameters are active per token. NVIDIA positions it for frontier reasoning, complex agentic workflows, long-context analysis, and agent orchestration rather than for high-volume cheap execution — which is the opposite emphasis from the smaller Lightning model in the same family.
The architecture matters practically because it constrains runtimes. A hybrid Mamba model needs a serving stack that supports it, and NVIDIA's own documentation names specific supported hardware and software versions. Confirming that before planning a deployment is cheaper than discovering it afterwards.
- Confirm your serving stack supports the hybrid Mamba architecture.
- Use active parameters as the cost and latency signal, not total parameters.
- Match the model to the step: orchestration and reasoning, not cheap high-volume execution.
Context at the model level, reasoning on a dial
NVIDIA documents a context length of up to 1 million tokens, though third-party indexes record smaller served windows on some providers — one recorded roughly 262,000 tokens. Separately, NVIDIA documents inference-time reasoning budget control and configurable reasoning mode through the chat template, which means the same deployment can be tuned per request.
That combination is the useful part for agent work: a long window for context and a reasoning dial for cost. As with any model, the served window and completion limit are provider-specific facts that must be checked against the endpoint you actually call rather than inferred from the model card.
- Record the context window and completion limit for your specific provider.
- Test at the context length you intend to use, not the maximum documented.
- Measure cost at each reasoning setting rather than assuming a default.
The hardware minimum is the real gate
NVIDIA's model card lists a minimum of eight H100-class GPUs, or four of certain Blackwell generations, for self-hosting. That is the single most important number in this article for most teams, because it settles the self-hosting question without further analysis: for anything smaller than a multi-GPU node, the hosted endpoint is the access route, not an alternative.
The weights are published under OpenMDW-1.1 and NVIDIA states the release is ready for commercial use. The permission exists; the hardware does not, for most teams. Reading the licence directly is still the right move if you plan to redistribute a modified model, since OpenMDW-1.1 is not the MIT licence and a summary of it is not the licence.
- Read the documented hardware minimum before planning self-hosting.
- Record which checkpoint and licence you adopted.
- Treat the hosted endpoint as the primary access route unless you operate the hardware.
A free endpoint on a frontier model is a research budget
Nemotron 3 Ultra is reachable at zero cost on NVIDIA's own endpoint and on at least one third-party gateway, where third-party tracking has recorded a 1-million-token context, roughly two-second first-token latency, around 28 tokens per second, and uptime well below a service level. The paid tier elsewhere is recorded at roughly fifty cents per million input and two dollars twenty per million output.
Those free-endpoint observations are point-in-time readings, not guarantees, and a hosted free tier on a frontier model carries the usual conditions: rate limits, concurrency caps, and no continuity commitment. For evaluation and prototyping that is genuinely valuable. For anything that must keep running, budget for the paid tier or design the fallback before the dependency.
- Record rate and concurrency limits before designing around a free endpoint.
- Measure latency and throughput yourself rather than relying on a third-party reading.
- Keep a paid fallback for anything that must keep running.