Cosmos3-Super-Reasoner: specs, benchmarks, and how to run it on Lyceum

Get started: call Cosmos3-Super-Reasoner on Lyceum

Integrating NVIDIA's Cosmos3-Super-Reasoner into your application requires zero new frameworks if you already use standard API clients. Lyceum provides a drop-in replacement for the OpenAI SDK, allowing you to route requests to secure European infrastructure by updating the base URL and API key. This approach can support GDPR compliance for your physical AI and video understanding workloads without requiring architectural rewrites, but the controller remains responsible for implementing and demonstrating compliance.

Below is the exact Python snippet to call the model using the standard OpenAI client.

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/Cosmos3-Super-Reasoner",
 messages=[{"role": "user", "content": "Hello!"}],
 max_tokens=256,
 )
 print(response.choices[0].message.content)

Pricing and region for Cosmos3-Super-Reasoner

When deploying physical AI models, predictable unit economics are critical. Cosmos3-Super-Reasoner is available on Lyceum's serverless inference API, priced per token. The model is priced at $0.10 per million input tokens and $0.30 per million output tokens.

All inference for this model runs in the eu-north1 region, so your video feeds, images, and proprietary text stay in the European Union. For enterprises building autonomous systems or analyzing factory floor footage, that removes the international transfer assessment that comes with routing data to US-based hyperscalers. You pay strictly per token with no base fees, allowing you to scale from prototyping to production efficiently.

What Cosmos3-Super-Reasoner is good at

Physical AI and omnimodal understanding

Cosmos3-Super-Reasoner is the reasoner tower of NVIDIA's Cosmos 3 Super, a 64B-parameter omnimodal model built from a 32B reasoner and a 32B generator. NVIDIA describes the architecture as a Mixture-of-Transformers with two towers: an autoregressive vision-language tower that interprets text, images and video, and a diffusion tower that generates video, audio and action outputs. The reasoner can be called on its own, which is what Lyceum serves, so requests take text, images and video and return text. That design allows the model to understand the physical world, making it highly effective for robotics, autonomous vehicles, and smart space environments. It can analyze complex scenes, track object permanence, and understand spatial relationships across video frames. For still images rather than sequences, we cover Wan Image and how to run it separately.

Complex multi-step reasoning

The model excels at breaking down real-world scenarios into structured state sequences. When analyzing fixed-camera footage from warehouses, transportation hubs, or factory assembly lines, Cosmos3-Super-Reasoner can reliably segment activity and reason about what is happening. It serves as a planning model, using prior knowledge and physics understanding to determine what steps an embodied agent should take next. This makes it ideal for generating action sequences and evaluating physical plausibility in simulated environments.

Video analytics and anomaly detection

For industrial vision applications, the model provides robust performance in detecting events and anomalies. It can process long video sequences to identify deviations from standard operating procedures on a manufacturing line or track specific activities in a logistics hub. By combining visual perception with deep reasoning capabilities, Cosmos3-Super-Reasoner allows engineering teams to build automated monitoring systems that understand context, rather than relying on brittle, hard-coded computer vision rules.

Limitations and trade-offs

Where Cosmos3-Super-Reasoner falls short

Cosmos3-Super-Reasoner is not the right tool for every workload. As a 32B-parameter model, it requires significant compute resources. Teams treating it as a lightweight drop-in replacement for smaller 8B models will encounter higher latency and increased token costs. For edge robotics or embedded hardware where millisecond latency is critical, the smaller Cosmos3-Nano variant is often a better choice.

Where accuracy degrades

NVIDIA's model card is explicit that the reasoning can be wrong: object states, causal relationships, spatial geometry, temporal ordering, agent intent, and future outcomes can all be misinferred, and long-context inputs can yield hallucinated entities or implausible predictions. Because the model carries no explicit physics simulator, object permanence, contact dynamics, and physical laws are approximated rather than computed, and quality degrades further in out-of-distribution environments and safety-critical edge cases. The technical report shows the same shape in numbers: on plain object counting, Cosmos 3 Super scores 89.1 on CountBenchQA against 93.6 for Qwen3-VL 32B, so a narrow counting or tracking model can still beat it on that one task.

Not a general-purpose text assistant

Cosmos3-Super-Reasoner is engineered for physical AI and video understanding. If your primary use case involves drafting emails, summarizing standard text documents, or generating code, this model is over-engineered and less efficient than dedicated text models like Llama 3 or Mistral. Its architecture is optimized for multimodal physical reasoning, and using it for pure text tasks wastes its omnimodal capabilities while incurring the overhead of a 32B-parameter vision-language model. Engineering teams should reserve this model for workloads that actually require physical world comprehension.

Benchmarks and how it compares

Cosmos3-Super-Reasoner benchmark results

NVIDIA evaluates the Cosmos 3 family across multiple benchmark suites targeting physical AI reasoning, generation quality, and domain-specific performance. Cosmos3-Super-Reasoner ranks at the top of its parameter class for understanding real-world environments.

Figures from Table 10 of the Cosmos 3 technical report, read 3 August 2026. The comparison column is the nearest open model in the same 32B block of that table.

Comparing to sibling models

Within the NVIDIA catalogue, the Cosmos 3 Super reasoner (32B) sits above the Cosmos 3 Nano reasoner (8B); they are the reasoning towers of a 64B and a 16B model respectively. NVIDIA positions Nano for workstation-grade compute and real-time robotics, and Super for datacenter deployment on Hopper and Blackwell GPUs, where it delivers the highest benchmark scores in the family.

Against generalist vision-language models, the advantage shows up in the physical AI domains rather than across the board. The technical report's own summary says Cosmos 3 is competitive with open models on general benchmarks while still trailing Gemini 3.1 Pro, and that in robotics, smart infrastructure, and driving it outperforms the open and closed models tested, with a small gap to Gemini 3.1 Pro in robotics. NVIDIA is equally clear that there is no physics simulator inside the model, so physical laws are approximated rather than enforced. However, this specialization means it requires more careful prompt engineering and region framing to extract structured state sequences effectively.

Using it in production

Production configuration for Cosmos3-Super-Reasoner

Deploying Cosmos3-Super-Reasoner effectively requires understanding its context limits and pricing structure. Lyceum lists the model with a 128K context window, enough to ingest long video sequences, high-resolution image batches, and extensive system prompts in a single API call. This deep context is essential for analyzing multi-minute fixed-camera footage or providing a robot with extensive historical state data before asking it to reason about its next action.

Lyceum serves this model on its serverless inference API, priced per token. The published price is $0.10 per million input tokens and $0.30 per million output tokens.

To understand the unit economics, consider a video analytics workload. If you pass a sequence of frames and text prompts totaling 50,000 input tokens, and the model generates a detailed 500-token structured JSON analysis of the physical events, the cost is minimal. The input costs $0.005, and the output costs $0.00015, resulting in a total API call cost of $0.00515.

All requests are routed to our eu-north1 region. For European manufacturing, logistics, and automotive companies, this ensures that proprietary factory footage and autonomous driving data remain strictly within the EU. You can scale your inference volume dynamically without committing to expensive reserved instances, paying only for the exact tokens processed during your physical AI evaluations.

Running Cosmos3-Super-Reasoner on EU-sovereign infrastructure

Why run Cosmos3-Super-Reasoner on Lyceum

Building physical AI systems requires processing highly sensitive data. Factory floor camera feeds, autonomous vehicle sensor logs, and proprietary robotics training data sent to US-based infrastructure may trigger compliance risks and require a case-by-case transfer assessment and, where necessary, supplementary measures. Lyceum provides an EU-native inference platform capable of serving heavy omnimodal models like Cosmos3-Super-Reasoner from European data centres. Learn more about GDPR-compliant LLM inference in Europe.

By hosting the model in the eu-north1 region, processing stays within European borders, with GDPR-compliant processing and no training on customer data. Lyceum runs European data centres in Spain, Paris and the Nordics, and serves this model from eu-north1, with serverless inference billed per token and GPU compute billed per second, and no base fee. You avoid the pain of managing your own hardware, dealing with cooling requirements, or fighting for GPU availability.

Furthermore, Lyceum provides open-stack transparency. We utilize optimized open-source inference engines like vLLM and NVIDIA Dynamo rather than locking you into a black-box proprietary stack. Our API is fully OpenAI-compatible, meaning your engineering team can switch from existing providers by changing a single URL string. You get the physical reasoning of NVIDIA's 32B Cosmos 3 reasoner, the scalability of serverless execution, and processing in European data centres (eu-north1 for this model), with no training on customer data and no data retained after processing, all without minimum commitments or egress fees.

Sources

[1] Develop Physical AI Reasoning, World, and Action Models with NVIDIA Cosmos 3; [2] Cosmos 3 - Research at NVIDIA; [3] nvidia/Cosmos3-Super - Hugging Face; [4] cosmos/inference_benchmarks.md at main - GitHub; [5] Cosmos 3: Omnimodal World Models for Physical AI (technical report)