NVIDIA: the Nemotron family of open models
Run NVIDIA Nemotronâ„¢ models built for complex agent workflows, engineered for efficient fine-tuning and inference on Crusoe infrastructure.
Deliver low latency where users are waiting
Real-time chat and voice apps live or die on time-to-first-token. Crusoe’s inference engine, powered by MemoryAlloy™ technology, caches context across the cluster and routes each request to the node best placed to answer it, so Nemotron models start responding faster when it matters most.
Serve text, vision, and speech on one platform
Nemotron models excel in a variety of agentic AI tasks. Crusoe hosts a wide range of them in Crusoe Intelligence Foundry so you can build pipelines that mix modalities without stitching together separate providers, keys, and billing.
Handle traffic spikes without overprovisioning
Serverless Inference scales Nemotron models automatically as demand rises and falls. You pay for the tokens you use, not for idle capacity sized for your busiest hour.
Fine-tune and deploy in Crusoe Intelligence Foundry
Frequently
asked questions
NVIDIA is the only model provider on Crusoe Cloud that also designs the GPUs its models run on, and it engineers its open Nemotron models for inference efficiency on that hardware. NVIDIA releases weights together with training data and recipes under the NVIDIA Open Model License and the Linux Foundation’s OpenMDW license, and in 2026 launched the NVIDIA Nemotron Coalition™ to advance open frontier models with other labs. On Crusoe Cloud, Nemotron models are available across Serverless Inference, Self-Serve Deployments, and Serverless Fine-Tuning, with support varying by model, running on Crusoe-operated infrastructure with no customer data sent to NVIDIA. Request a Tailored Deployment for bespoke optimization and a dedicated endpoint.
Nemotron models are open-weight models released under permissive licenses, such as the NVIDIA Open Model License and the Linux Foundation’s OpenMDW license. Both permit commercial use, modification, and redistribution, subject to their terms, including NVIDIA's acceptable use terms. Neither is OSI-approved, and terms vary by release,so check each model card before deploying.
Choose based on workload size and input type. Smaller Nemotron models activate only a few billion parameters per token, which makes them fast and cost-efficient for high-volume tasks and agents. Larger Nemotron models deliver stronger reasoning and coding for complex, multi-step work. For audio, image, video, or voice applications, choose one of the Nemotron models specialized for multimodal or speech.
Yes, select Nemotron models are multimodal. They accept combinations of text, image, video, and audio input for tasks like document understanding, video analysis, and transcription. NVIDIA also offers end-to-end speech models built for real-time voice applications. Supported input types vary by model, so check each model’s details.
A hybrid Mamba-Transformer architecture combines Mamba state-space layers with standard attention layers. Mamba layers process sequences with memory that stays constant as context grows, unlike attention, whose KV cache grows with every token. A smaller number of attention layers preserves precise recall. Recent Nemotron models pair this design with mixture-of-experts routing for higher throughput and efficient long-context inference.
Yes, select Nemotron models are available for Serverless Fine-Tuning. Upload your dataset, run a LoRA job with pre-configured defaults, then deploy the resulting checkpoint on a dedicated endpoint through Self-Serve Deployments, or download the weights. Each model’s details show whether it currently supports fine-tuning.
%204.png)