Nemotron-3-Ultra-550b: specs, benchmarks, and how to run it on Lyceum

Get started: call Nemotron-3-Ultra-550b on Lyceum

To begin building with Nemotron-3-Ultra-550b, you can use the standard OpenAI Python SDK. Because Lyceum Technology provides an OpenAI-compatible API, switching your existing applications requires zero code changes beyond updating the base URL and your API key. The model string for this endpoint is nvidia/Nemotron-3-Ultra-550b-a55b.

from openai import OpenAI
 
 client = OpenAI(
 base_url="<the base URL shown in your Lyceum dashboard>",
 api_key="<your lyceum api key>",
 )
 response = client.chat.completions.create(
 model="nvidia/Nemotron-3-Ultra-550b-a55b",
 messages=[{"role": "user", "content": "Hello!"}],
 max_tokens=256,
 )
 print(response.choices[0].message.content)

Pricing and region for Nemotron-3-Ultra-550b

This model is available through Lyceum's Serverless Inference, an OpenAI-compatible endpoint suited to demanding agentic workloads. This model is globally hosted rather than EU-hosted, one of four such models in the catalogue, and it never receives your traffic unless you explicitly select it. Pricing is strictly pay-per-token, with no minimum commitments or base fees. The input cost is $1.00 per million tokens, and the output cost is $3.00 per million tokens.

For teams transitioning off hyperscaler credits, this pay-as-you-go structure ensures you only pay for the exact compute you consume. If you need predictable throughput at scale, you can also burst this model onto dedicated GPU virtual machines billed per second. By utilizing our serverless GPU inference platform, your engineering team can avoid the operational overhead of managing complex Kubernetes clusters or dealing with out-of-memory errors on massive 550B parameter models.

What Nemotron-3-Ultra-550b is good at

Built for long-running autonomous agents

Nemotron-3-Ultra-550b is engineered specifically for agentic workflows that require sustained reasoning over extended periods. On Lyceum the model is served with a 128,000-token context window, enough to preserve agent state, system logs, and execution plans across sustained sessions. NVIDIA reports 95% for the model on the RULER long-context benchmark at its native 1M-token limit, the highest score in NVIDIA's published comparison table.

Hybrid Mamba-Transformer architecture

The model utilizes a Latent Mixture-of-Experts (MoE) architecture that interleaves Mamba-2 state space layers with traditional Transformer attention layers. The Mamba layers provide linear scaling and sequence efficiency for massive contexts, while the Transformer layers deliver the precision reasoning required for complex logic. This hybrid approach allows the model to activate only 55 billion parameters during inference while leveraging the knowledge capacity of its 550 billion total parameters.

High-throughput speculative decoding

Nemotron-3-Ultra-550b includes Multi-Token Prediction (MTP) layers that enable native speculative decoding. This architectural choice significantly accelerates generation speeds. NVIDIA reports up to 5.9x higher inference throughput than GLM-5.1 and 4.8x higher than Kimi K2.6 at the 8k and 64k token settings (NVIDIA Research), making it efficient for production deployments where time-to-first-token and overall generation speed are critical.

Granular reasoning control

For complex tasks, the model supports inference-time reasoning budget control. Developers can utilize parameters like enable_thinking and reasoning_budget to dictate how much compute the model allocates to generating internal reasoning traces before it outputs a final answer.

Limitations and trade-offs

Where Nemotron-3-Ultra-550b falls short

While Nemotron-3-Ultra-550b excels at agentic orchestration, it requires careful resource management and has specific performance trade-offs compared to other frontier models.

First, the hardware requirements for self-hosting are massive. With 550 billion total parameters, the BF16 weights alone consume over 1.1 TB of storage. Running this model locally or on dedicated infrastructure requires a multi-node cluster of high-end GPUs with tensor parallelism, since 1.1 TB of BF16 weights plus KV cache does not fit in a single 8x H200 node. For teams without access to large-scale clusters, using a managed API is the only practical way to access the model.

Second, while the model ties or beats competitors in general agent productivity, it trails slightly in raw coding benchmarks. On Terminal-Bench 2.0, Nemotron-3-Ultra-550b scores 54%, which is lower than Kimi K2.6 at 67% and GLM 5.1 at 64%. Teams building dedicated coding assistants might find specialized coding models more effective for raw syntax generation.

Finally, the model's advanced reasoning capabilities are token-intensive. When the thinking mode is enabled, the model generates extensive internal reasoning traces before providing the final answer. Because these reasoning steps consume output tokens, the cost per request can increase significantly if the reasoning budget is not strictly capped. Developers must actively manage the reasoning_budget parameter to prevent cost overruns on simpler queries that do not require deep logical deduction.

Benchmarks and how it compares

Nemotron-3-Ultra-550b benchmark results

NVIDIA has published extensive benchmark data demonstrating how Nemotron-3-Ultra-550b performs against other frontier-class open models. The model is particularly strong in agent productivity and instruction following, though it faces stiff competition in raw coding tasks.

Source: NVIDIA Technical Blog, Table 1. NVIDIA reports SWE-bench Verified as a range because the score moves with the agent scaffold, from Pi and OpenHands through to Mini SWE Agent. A dash marks a benchmark NVIDIA does not publish for that model.

When compared to its sibling model, Nemotron-3-Super-120B, the Ultra variant offers significantly higher capacity for complex, multi-step reasoning. While the 120B Super model activates only 12B parameters and is optimized for maximum compute efficiency, the 550B Ultra model activates 55B parameters, providing the deep knowledge retrieval and logical rigor required for enterprise-grade research and autonomous agent orchestration. For tasks requiring the absolute highest accuracy and longest context retention, Nemotron-3-Ultra-550b is the superior choice within the NVIDIA catalogue.

Using it in production

Production configuration for Nemotron-3-Ultra-550b

Deploying Nemotron-3-Ultra-550b in production requires understanding its region and pricing structure. On Lyceum Technology, this model is served through Serverless Inference, which is billed per token and has no service tiers or reserved capacity classes.

The endpoint for nvidia/Nemotron-3-Ultra-550b-a55b is globally hosted rather than EU-hosted, and it never receives your traffic unless you explicitly select it. When configuring your API requests, you can use the model's 128,000-token context window, enough to pass large files, documentation sets, or long conversation histories in a single prompt.

Pricing is calculated strictly per token. The input rate is $1.00 per million tokens, and the output rate is $3.00 per million tokens. To understand the unit economics, consider a deep research task where you submit 100,000 tokens of source material and the model generates a 2,000-token analysis, including its internal reasoning trace. The input cost would be $0.10, and the output cost would be $0.006, resulting in a total API call cost of $0.106.

Because the model supports reasoning traces, you should configure your max_tokens parameter generously to ensure the model has enough output space to complete its thought process. We recommend enabling streaming in your API calls so your application can process the reasoning trace in real time, reducing perceived latency for end users while the model computes its final answer.

Why run Nemotron-3-Ultra-550b on Lyceum

Why run Nemotron-3-Ultra-550b on Lyceum

Choosing the right infrastructure provider is critical when deploying frontier models like Nemotron-3-Ultra-550b. Lyceum Technology pairs high performance with a transparent, developer-first platform.

This model is globally hosted rather than EU-hosted, and it never receives your traffic unless you explicitly select it. Integration is straightforward: because our Serverless Inference API is fully OpenAI-compatible, your engineering team can switch to Lyceum by updating a single base URL, with no SDK rewrites or proprietary client libraries.

Pricing is strictly pay-per-token with no idle time and no base fees, so you only pay for the compute you actually consume. Our platform is built on open-stack transparency: we run optimized open inference engines like vLLM and NVIDIA Dynamo rather than black-box proprietary stacks, which guarantees customer portability by design. Billing is unified across serverless and dedicated GPU usage, and there are zero egress fees. Another reasoning-focused model is covered in our notes on Cosmos3-Super-Reasoner and how to run it.

When you need predictable throughput at scale, you can burst from the pay-per-token API onto dedicated GPU virtual machines billed per second, for example a dedicated 8x H200 cluster, without changing your application code. The token economics also change, since open-weight models are served at a fraction of the per-token rate charged for closed frontier models. Current Lyceum rates are published on our pricing page. Learn more about how this works in our guide to serverless GPU inference.

Sources

[1] NVIDIA Technical Blog; [2] Hugging Face Model Card; [3] NVIDIA Technical Report; [4] NVIDIA Build