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

Get started: call Nemotron-3-Nano-Omni on Lyceum

Integrate Nemotron-3-Nano-Omni into your agentic workflows using the standard OpenAI Python SDK. Because Lyceum Technology provides an OpenAI-compatible API, switching providers requires zero code changes to your application logic. Update the base URL and provide your Lyceum API key to route multimodal requests to our secure European infrastructure. This drop-in compatibility accelerates deployment for engineering teams migrating away from hyperscaler environments.

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-Nano-Omni",
 messages=[{"role": "user", "content": "Hello!"}],
 max_tokens=256,
 )
 print(response.choices[0].message.content)

Pricing and region for Nemotron-3-Nano-Omni

This model is served through Lyceum Serverless Inference, which is priced for cost-efficient inference on high-volume agentic tasks and perception workloads. The pricing is $0.06 per million input tokens and $0.24 per million output tokens. All inference for this endpoint is processed in the eu-north1 region, ensuring your multimodal data, whether sensitive documents, audio recordings, or video frames, is processed in European data centres. This setup is ideal for teams building complex AI agents that require rapid perception across multiple modalities while maintaining strict data residency and compliance. Serverless Inference bills per token with no base fee, so you pay for the tokens your application actually sends and receives.

What Nemotron-3-Nano-Omni is good at

Unified omni-modal reasoning

Nemotron-3-Nano-Omni excels at processing multiple data types simultaneously without relying on fragmented model pipelines. Traditional agentic systems often stitch together separate vision, speech, and language models, which increases inference hops and orchestration complexity. NVIDIA designed this model to handle video, audio, image, and text inputs natively within a single perception-to-action loop. It utilizes the CRADIOv4 encoder for high-resolution vision tasks like optical character recognition (OCR) and document parsing. For audio, the Parakeet encoder processes transcription, spoken queries, and environmental sounds. Video reasoning is accelerated by 3D convolutional layers and Efficient Video Sampling (EVS), allowing the model to analyze temporal-spatial data across long clips.

High-efficiency MoE architecture

The model achieves high throughput through a hybrid Mamba-Transformer Mixture-of-Experts (MoE) architecture. While it contains 30 billion total parameters, it activates only about 3 billion parameters per token during inference. This selective activation allows Nemotron-3-Nano-Omni to operate with the speed and cost profile of a small dense model while delivering the reasoning capabilities of a much larger system. The inclusion of Mamba selective state-space layers provides efficient long-context processing, and NVIDIA's model card lists a maximum context of 256K tokens; Lyceum serves the model with a 128K window. This makes the model highly effective for analyzing multi-hour audio recordings, extended screen sessions, and dense enterprise documents without dropping critical context mid-task. NVIDIA reports that for video reasoning at a fixed interactivity threshold, this efficiency translates into up to 9.2x greater effective system capacity than alternative open omni models.

Limitations and trade-offs

Where Nemotron-3-Nano-Omni falls short

While Nemotron-3-Nano-Omni is a highly capable perception engine, it is not designed to replace massive frontier models for deep logical reasoning or complex software engineering tasks. Because it activates only 3 billion parameters per token, its pure text-based reasoning depth cannot match 100B+ parameter models like Nemotron-3-Super or Nemotron-3-Ultra. If your workload requires advanced mathematical proofs, intricate coding architectures, or deep strategic planning, a larger text-specialized model will yield better results.

Another trade-off involves output formatting and timestamp generation during long video or audio analysis. When processing extensive multimedia files, the model can occasionally struggle with precise timestamp alignment or enter repetitive loops if the output formatting instructions are too complex. Engineers deploying this model for video action recognition or audio lyric extraction should enforce strict max_tokens limits to prevent runaway responses and budget overruns.

Furthermore, the model is optimized as a sub-agent within a broader multi-agent system. It is intended to act as the "eyes and ears" of an AI application, passing its structured observations to a heavier reasoning model for final decision-making. Relying on Nemotron-3-Nano-Omni as a standalone system for highly complex, multi-step logical deductions will likely expose the limitations of its smaller active parameter count. Teams should architect their systems to utilize this model for rapid multimodal perception rather than heavy cognitive lifting.

Benchmarks and how it compares

Nemotron-3-Nano-Omni benchmark results

Nemotron-3-Nano-Omni demonstrates strong performance across multimodal evaluations, particularly in document intelligence and video understanding. It competes favorably against both open-weight and proprietary models in its parameter class.

Source: the NVIDIA Nemotron-3-Nano-Omni-30B-A3B-Reasoning model card, benchmark scores table, read 3 August 2026.

When evaluated on MMLongBench-Doc, a rigorous test for long-context document understanding, Nemotron-3-Nano-Omni scores 57.5, against 38 for the previous Nemotron Nano VL V2. This highlights the effectiveness of the CRADIOv4 vision encoder for complex enterprise documents.

In video reasoning, the model achieves a 72.2 on Video-MME, and 55.4 on WorldSense. The integration of 3D convolutional layers and Efficient Video Sampling gives NVIDIA's model a distinct advantage in temporal-spatial tasks. For audio-visual cross-modal reasoning, it scores 74.5 on the DailyOmni benchmark. Nemotron-3-Nano-Omni activates only 3 billion parameters per token to deliver these benchmark results with significantly higher throughput and lower inference costs than dense models, making it a highly efficient choice for production multimodal pipelines. A larger vision-language model is covered in our notes on Qwen2.5-VL-72B and how to run it.

Using it in production

Production configuration for Nemotron-3-Nano-Omni

Deploying Nemotron-3-Nano-Omni effectively requires understanding its context window and pricing. Lyceum serves the model with a 128,000-token context window, which is useful for processing long-form media such as multi-hour meeting recordings, extensive PDF documents, or continuous video streams. Serverless Inference has no service tiers, and Nemotron-3-Nano-Omni is built for high-throughput, latency-sensitive applications where rapid perception is critical.

When configuring your API requests, you should carefully manage the max_tokens parameter. Because multimodal models can occasionally generate verbose descriptions of visual or audio inputs, setting a strict output limit prevents unnecessary token consumption. For example, if you are extracting specific data points from a video, capping the output at 500 tokens ensures concise responses and predictable costs.

The per-token pricing model makes this highly cost-effective for variable workloads. At $0.06 per million input tokens and $0.24 per million output tokens, processing a large batch of documents is highly economical. For instance, analyzing a dataset that consumes 5 million input tokens and generates 500,000 output tokens would cost exactly $0.30 for the input and $0.12 for the output, totaling $0.42. All requests are routed through the eu-north1 region, ensuring that your production data is processed on secure European infrastructure with zero egress fees.

Running Nemotron-3-Nano-Omni on EU-sovereign infrastructure

Why run Nemotron-3-Nano-Omni on Lyceum

For European AI teams and enterprises handling sensitive multimodal data, data residency is a strict requirement. Running Nemotron-3-Nano-Omni on Lyceum Technology means your audio recordings, video files, and proprietary documents are processed in European data centres. All inference for this model occurs in the eu-north1 region, with no training on customer data and no data retained after processing. Unlike US-based API providers that route traffic through overseas data centers, Lyceum serves this model from European data centres in Spain, Paris and the Nordics.

That European footprint lets us offer competitive per-token pricing for teams that need their multimodal data processed in Europe. Use our serverless inference engine to benefit from per-token billing with no base fee. You pay strictly for the tokens you consume, eliminating the need to maintain expensive, idle GPU instances for bursty multimodal workloads.

Furthermore, Lyceum Technology prioritizes open-stack transparency. We utilize optimized inference engines like vLLM and NVIDIA Dynamo rather than locking you into proprietary, black-box architectures. This ensures complete customer portability and predictable performance. If you are migrating from a hyperscaler environment, our OpenAI-compatible endpoints allow you to transition your GDPR-compliant LLM inference to Lyceum in minutes, securing your data sovereignty while significantly reducing your infrastructure costs.

Sources

[1] Nemotron 3 Nano Omni: Efficient and Open Multimodal Intelligence; [2] nvidia/Nemotron-3-Nano-Omni-30B-A3B-Reasoning-BF16 - Hugging Face; [3] NVIDIA Nemotron 3 Nano Omni Powers Multimodal Agent Reasoning; [4] NVIDIA Technical Blog