Nemotron-3-Nano-30B: specs, benchmarks, and how to run it on Lyceum

Get started: call Nemotron-3-Nano-30B on Lyceum

Deploying NVIDIA's Nemotron-3-Nano-30B on Lyceum requires zero new frameworks. Because our Serverless Inference API is fully OpenAI-compatible, you can route your existing application traffic to this highly efficient MoE model by updating your base URL and API key. This eliminates the friction of learning proprietary SDKs and allows your engineering team to focus on building features rather than managing infrastructure. The larger sibling is documented separately in our notes on Nemotron-3-Super-120b-a12b and how to run it.

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/NVIDIA-Nemotron-3-Nano-30B-A3B",
 messages=[{"role": "user", "content": "Hello!"}],
 max_tokens=256,
 )
 print(response.choices[0].message.content)

Pricing and region for Nemotron-3-Nano-30B

Lyceum Technology serves this model through Serverless Inference, priced for cost-efficient, high-throughput workloads. The pricing is $0.06 per million input tokens and $0.24 per million output tokens. All inference for this endpoint runs on our EU-sovereign infrastructure in the eu-north1 region, so prompts and completions are processed in European data centres.

This drop-in compatibility means your engineering team can evaluate Nemotron-3-Nano-30B against your current production models in minutes. There are no minimum commitments, no subscription tiers, and no base fees - you pay strictly per token processed. For teams migrating from hyperscaler environments, this provides immediate cost visibility and eliminates the need to provision dedicated GPU instances for bursty agentic workloads.

What Nemotron-3-Nano-30B is good at

Agentic reasoning and tool use

Nemotron-3-Nano-30B was engineered specifically for multi-step agentic workflows. NVIDIA trained the model to generate an internal reasoning trace before outputting a final response. This "thinking" capability allows the model to map out logic puzzles, execute complex routing decisions, and handle multi-constraint instructions with high reliability. It excels at structuring messy text into clean JSON and utilizing external tools, making it highly effective as a sub-agent in larger AI systems.

High-throughput efficiency

The architecture of Nemotron-3-Nano-30B is a hybrid Mamba-Transformer Mixture-of-Experts (MoE). While the model contains 30 billion parameters in total, it only activates approximately 3 billion parameters (hence the "A3B" designation) per token during inference. By combining 23 Mamba-2 layers with sparse MoE layers and grouped-query attention, the model achieves exceptional throughput and low latency. This sparse activation makes it highly cost-effective for high-volume tasks like log analysis or continuous data processing.

Long-context processing

Traditional transformer models suffer from severe memory degradation as context length increases. By utilizing Mamba-2 state-space layers, Nemotron-3-Nano-30B efficiently handles long inputs without the typical memory blowup. NVIDIA's model card publishes context support up to 1M tokens, with the Hugging Face default set to 256K; Lyceum serves the model with a 128,000-token context window. This allows developers to feed entire codebases, extensive documentation, or long conversation histories into the prompt. The hybrid architecture ensures that the model maintains precision reasoning over these long horizons, which is critical for retrieval-augmented generation (RAG) pipelines and persistent-memory agents.

Limitations and trade-offs

Where Nemotron-3-Nano-30B falls short

While Nemotron-3-Nano-30B delivers exceptional performance for logic and routing, it is not a general-purpose model for every workload. Its highly specialized training makes it less ideal for creative writing, open-ended brainstorming, or nuanced roleplay. Teams building creative applications will find that the model's responses lean toward structured, analytical formats rather than natural, conversational prose.

Additionally, the model's high accuracy is heavily dependent on its reasoning trace. While developers can configure the chat template to force the model to output a final answer without generating intermediate "thinking" tokens, doing so results in a measurable decrease in accuracy on complex prompts. If your application requires instantaneous, zero-latency responses without the overhead of reasoning tokens, a dense model of a similar size might provide more consistent zero-shot performance.

Finally, as a 30B parameter model, it lacks the broad, encyclopedic world knowledge found in frontier models exceeding 100B parameters. If a query relies on obscure facts rather than provided context, the sparse expert activation may result in generalized or flat responses. It performs best when supplied with explicit context via RAG rather than relying on its internal weights for factual recall.

Benchmarks and how it compares

Nemotron-3-Nano-30B benchmark results

NVIDIA evaluated Nemotron-3-Nano-30B against leading open-weight models in its size class, demonstrating significant advantages in mathematical reasoning and coding tasks. The hybrid MoE architecture proves highly competitive against dense models.

Source: NVIDIA's Nemotron-3-Nano-30B-A3B-BF16 model card, read 3 August 2026.

In the mathematical domain, Nemotron-3-Nano-30B achieves 89.1% on AIME 2025 without tool assistance, outperforming Qwen3-30B-A3B. When Python tools are enabled, its score rises to 99.2%, highlighting its strength in agentic workflows. On software engineering tasks like LiveCodeBench v6, it records 68.3%, ahead of Qwen3-30B-A3B at 66.0% and GPT-OSS-20B at 61.0%. The Arena-Hard-v2 score of 67.7% further validates its reliability in complex, multi-step instructions, making it a robust choice for developers building autonomous agents.

Using it in production

Production configuration for Nemotron-3-Nano-30B

When deploying Nemotron-3-Nano-30B in production, understanding its reasoning budget is critical. Because the model generates a reasoning trace before its final answer, output token counts will be higher than those of standard non-reasoning models. You can control this by adjusting the reasoning budget parameters in your API request, allowing you to cap the "thinking" tokens to keep inference costs predictable.

Lyceum Technology serves this model through Serverless Inference, which is priced for cost-efficiency and high throughput. Hosted in our eu-north1 region, it leverages our open-stack infrastructure - including vLLM and TensorRT-LLM - to maximize the hardware-aware efficiency of the model's NVFP4 quantization.

To calculate production costs, consider a log analysis agent processing server errors. If the agent receives a 15,000-token input prompt (context) and generates a 500-token reasoning trace followed by a 200-token JSON output (700 output tokens total), the cost math is straightforward. At $0.06 per million input tokens, the prompt costs $0.0009. At $0.24 per million output tokens, the generation costs $0.000168. The total cost for this complex, reasoning-heavy task is approximately $0.001 per request. This aggressive pricing allows engineering teams to scale multi-agent systems without the prohibitive costs associated with hyperscaler APIs.

Running Nemotron-3-Nano-30B on EU-sovereign infrastructure

Why run Nemotron-3-Nano-30B on Lyceum

For European AI teams, data sovereignty is a hard requirement, not an optional feature. Running Nemotron-3-Nano-30B on Lyceum Technology keeps your inference workloads in European data centres. Our eu-north1 region provides GDPR-compliant processing, so sensitive data - whether medical records, financial logs, or proprietary code - is not routed into US jurisdictions. Lyceum serves this endpoint from European data centres in Spain, Paris and the Nordics, and bills strictly per token with no base fee.

Migrating to Lyceum requires minimal engineering effort. Because our Serverless Inference API is fully OpenAI-compatible, you can switch providers by updating a single base URL. There is no need to rewrite your application logic or learn proprietary SDKs. You benefit from open-stack transparency, avoiding the vendor lock-in associated with black-box proprietary engines.

Furthermore, our pay-per-token billing model eliminates the financial risk of idle compute. You scale to zero when traffic drops and pay only for the exact tokens processed. For teams transitioning off expiring hyperscaler credits, Lyceum offers a sustainable, high-performance path forward. Whether you are deploying serverless GPU inference for the first time or scaling a massive multi-agent architecture, Lyceum provides the secure, sovereign foundation your engineering team requires.

Sources

[1] NVIDIA Nemotron-3-Nano-30B-A3B-BF16 Model Card; [2] Nemotron 3 Nano: Open, Efficient Mixture-of-Experts Hybrid Mamba-Transformer Model; [3] NVIDIA Developer: Inside NVIDIA Nemotron 3, Techniques, Tools, and Data That Make It Efficient and Accurate