Nemotron-Ultra-253B: specs, benchmarks, and how to run it on Lyceum

Get started: call Nemotron-Ultra-253B on Lyceum

You can access Nemotron-Ultra-253B through Lyceum Technology's Serverless Inference API. Because our API is fully OpenAI-compatible, you can switch to our EU-hosted infrastructure by updating your base URL and API key.

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/Llama-3_1-Nemotron-Ultra-253B-v1",
 messages=[{"role": "user", "content": "Hello!"}],
 max_tokens=256,
 )
 print(response.choices[0].message.content)

Pricing and region for Nemotron-Ultra-253B

Lyceum Technology serves this model through Serverless Inference, billed per token with no base fee. There are no service tiers to pick: the model you choose sets the balance between latency and capability, and Nemotron-Ultra-253B is the high-capability reasoning option. The model is hosted in the eu-north1 region, so requests to it are processed in European data centres.

• Input pricing: $0.60 per million tokens
• Output pricing: $1.80 per million tokens
• Billing: per token, no base fee
• Region: eu-north1

What Nemotron-Ultra-253B is good at

Efficient frontier-level reasoning

Nemotron-Ultra-253B was built to solve a specific infrastructure problem: running a frontier-class reasoning model without requiring a massive GPU cluster. By applying Neural Architecture Search (NAS) and vertical compression to the Llama 3.1 405B architecture, NVIDIA reduced the parameter count to 253B. This allows the model to fit entirely on a single 8xH100 node for inference, significantly lowering the hardware barrier while maintaining top-tier reasoning capabilities for enterprise deployments.

Dual-mode chain-of-thought

Unlike models that force reasoning on every prompt, Nemotron-Ultra-253B features a dual-mode operation controlled via the system prompt. By including "detailed thinking on" or "detailed thinking off" in the system message, developers can toggle the model's chain-of-thought generation. This flexibility means you can use the same model for complex, multi-step math problems and standard, low-latency chat interactions without wasting output tokens on unnecessary reasoning.

Tool calling and RAG

The model underwent extensive post-training using Group Relative Policy Optimization (GRPO) specifically targeted at tool calling and Retrieval-Augmented Generation (RAG). It reliably outputs structured JSON, follows complex multi-step instructions, and manages context effectively across its 128K token window, making it an excellent engine for agentic workflows that require both deep analysis and predictable formatting. For teams building autonomous agents, this combination of structured output reliability and deep reasoning makes it a highly capable core engine. The model can analyze a user request, determine which external tools to call, and synthesize the results into a coherent final answer.

Limitations and trade-offs

Where Nemotron-Ultra-253B falls short

While Nemotron-Ultra-253B is highly optimized, it is still a massive 253-billion-parameter dense model. Unlike Mixture-of-Experts (MoE) architectures that only activate a fraction of their parameters per token, this model activates all 253B parameters during every forward pass. This dense architecture requires substantial VRAM and compute, meaning it cannot be run on smaller hardware setups like a single GPU or consumer-grade hardware. Teams looking for local deployment will need enterprise-grade infrastructure.

Additionally, the model's reasoning mode generates significant Key-Value (KV) cache pressure. When "detailed thinking on" is enabled, the model produces long reasoning chains before delivering the final answer. This consumes a large portion of the 128K context window and reduces overall throughput for concurrent requests. Teams deploying this model for high-concurrency applications must carefully manage KV cache allocation and consider whether the reasoning mode is strictly necessary for every query.

Finally, while it excels in scientific reasoning and math, it is not a multimodal model. Teams requiring native image or video processing will need to pair it with a dedicated vision model or look toward multimodal alternatives. It is strictly a text-in, text-out reasoning engine. Furthermore, because the reasoning toggle relies on exact system prompt phrasing, developers must ensure their prompt management systems correctly inject the required strings. Failing to include the exact "detailed thinking on" phrase will result in standard generation, which may lead to lower accuracy on complex logic puzzles. The current generation of that family is covered in our notes on Nemotron-3-Ultra-550b and how to run it.

Benchmarks and how it compares

Nemotron-Ultra-253B benchmark results

NVIDIA's dual-mode approach allows the model to scale its intelligence at inference time. Enabling reasoning mode yields massive performance jumps across complex benchmarks, proving the efficacy of the GRPO post-training phase.

Source: the NVIDIA Llama-3.1-Nemotron-Ultra-253B-v1 model card, reasoning-off and reasoning-on columns, read 3 August 2026.

Comparison to sibling models

When compared to the original Meta Llama 3.1 405B, Nemotron-Ultra-253B retains the vast majority of its intelligence while requiring roughly 40 percent less VRAM. This makes it a far more practical choice for teams transitioning off hyperscaler credits who need to manage infrastructure costs.

Against DeepSeek R1, Nemotron-Ultra-253B actually wins on GPQA Diamond and LiveCodeBench. While DeepSeek R1 holds a slight edge in MATH-500 (97.3% vs 97.0%), Nemotron's dense architecture avoids the complex MoE routing overhead and fits on a single 8xH100 node. DeepSeek R1's 671B total parameter count typically requires a 16-GPU cluster for BF16 inference, making Nemotron-Ultra-253B a much more accessible option for self-hosting or dedicated cloud deployments. For teams prioritizing coding and scientific reasoning, Nemotron offers a superior balance of intelligence and hardware efficiency.

Using it in production

Production configuration for Nemotron-Ultra-253B

When deploying Nemotron-Ultra-253B via Lyceum Technology's Serverless Inference product, requests run in the eu-north1 region. There are no tiers to configure: this model is built for complex reasoning and accuracy rather than raw speed, so pick a smaller model from the catalogue when time-to-first-token matters more.

To control the model's reasoning behavior, you must configure the system prompt. Injecting "detailed thinking on" instructs the model to generate a chain-of-thought before answering, which is ideal for coding and math. If you need lower latency for standard chat, use "detailed thinking off". Because the model supports a 128K context window, you can safely pass large documents for RAG workflows, but be mindful of the output token consumption when reasoning is enabled.

Calculating per-token pricing

Lyceum Technology charges $0.60 per million input tokens and $1.80 per million output tokens for this model. Consider a RAG application that processes a 10,000-token document and generates a 1,500-token reasoned response.

• Input cost: 10,000 tokens * ($0.60 / 1,000,000) = $0.006
• Output cost: 1,500 tokens * ($1.80 / 1,000,000) = $0.0027
• Total cost per request: $0.0087

This pay-per-token model allows you to scale from zero without committing to the massive upfront cost of an 8xH100 cluster. You only pay for the exact compute your application requires, making it highly cost-effective for bursty or unpredictable workloads.

Running Nemotron-Ultra-253B on EU-sovereign infrastructure

Why run Nemotron-Ultra-253B on Lyceum

For European enterprises, deploying a 253B-parameter model typically means relying on US-based hyperscalers or API providers, which introduces significant data privacy risks. Lyceum Technology solves this by offering Nemotron-Ultra-253B on our fully EU-sovereign infrastructure. Hosted in our eu-north1 region, your inference workloads are processed entirely within European borders, which means GDPR-compliant processing in European data centres.

Lyceum serves Nemotron-Ultra-253B from European data centres in Spain, Paris and the Nordics, with GPU compute billed per second and no base fee. Being a focused GPU provider rather than a reseller sitting on top of another cloud is what keeps the per-token rates competitive. Furthermore, our open-stack transparency, powered by vLLM and NVIDIA Dynamo, ensures you are never locked into a proprietary black-box inference engine. Because Lyceum does not charge egress fees, you can move your data and model outputs without the hidden costs associated with traditional cloud providers. You retain full visibility into how your workloads are executed.

Because our Serverless Inference API is a drop-in replacement for the OpenAI SDK, your engineering team can migrate to Lyceum in minutes. You get an NVIDIA reasoning model, the reliability of a managed API, and inference that stays in European data centres. Whether you are building autonomous agents or complex RAG pipelines, Lyceum provides the secure foundation your enterprise requires. By combining NVIDIA's highly optimized model architecture with Lyceum's purpose-built European cloud, AI teams can finally achieve frontier-level performance without compromising on data privacy or infrastructure costs.

Sources

[1] NVIDIA Llama-3.1-Nemotron-Ultra-253B-v1 Model Card; [2] arXiv: Llama-Nemotron, Efficient Reasoning Models; [3] NVIDIA Developer: Build Enterprise AI Agents with Advanced Open NVIDIA Llama Nemotron Reasoning Models