NVIDIA Nemotron 3 Ultra launches on AWS SageMaker with 550B parameters, 1M token context window

NVIDIA Nemotron 3 Ultra is now available on Amazon SageMaker JumpStart with 550 billion total parameters and 55 billion active parameters. The model uses a hybrid Transformer-Mamba Mixture-of-Experts (MoE) architecture and supports context windows up to 1 million tokens.

Model specifications

• Architecture: Hybrid Transformer-Mamba MoE
• Parameters: 550B total / 55B active per forward pass
• Context window: 1 million tokens
• Precision: NVFP4 format
• Modality: Text-to-text

The MoE architecture activates only 55 billion of the 550 billion total parameters per inference pass. According to NVIDIA, this design delivers 5x faster inference and up to 30% lower cost for agentic workloads compared to dense models of equivalent quality.

Deployment and pricing

Nemotron 3 Ultra deploys via one-click on SageMaker JumpStart using GPU instances including ml.p5en.48xlarge, ml.p5.48xlarge, or ml.g7e.48xlarge. AWS notes that these GPU instances cost several dollars per hour while running. Specific per-token pricing has not been disclosed.

The model is optimized for the NVFP4 format, a precision type designed to reduce hosting costs and improve inference speed.

Target use cases

NVIDIA positions Nemotron 3 Ultra specifically for multi-turn agentic workflows that span hundreds of interaction turns:

• Agent orchestration systems that coordinate multiple sub-agents
• Coding agents that generate, test, debug, and iterate on code across large repositories
• Research synthesis tasks requiring extended context coherence
• Multi-step enterprise automation with decision branching

The million-token context window allows agents to maintain state across extended tool-calling chains and planning loops.

Technical implementation

The hybrid Transformer-Mamba architecture combines traditional Transformer attention mechanisms with Mamba's structured state-space models. This architectural choice aims to maintain throughput at extended context lengths while keeping compute costs lower than dense models.

Developers can deploy using SageMaker Studio's interface or the SageMaker Python SDK. The model accepts standard chat completion payloads with configurable max_tokens, temperature, and top_p parameters.

Availability

Nemotron 3 Ultra is available immediately on Amazon SageMaker JumpStart. The model is described as "open" though specific licensing terms were not detailed in the announcement.

What this means

Nemotron 3 Ultra represents NVIDIA's direct entry into models purpose-built for agentic AI workflows. The 10:1 ratio between total and active parameters through MoE, combined with the 1M token context window, directly addresses the sustained compute demands of multi-turn agent interactions. The NVFP4 format optimization suggests NVIDIA is leveraging hardware-specific acceleration unavailable to other model providers. However, without independent benchmarks or disclosed per-token pricing, comparisons to existing agent-optimized models like Anthropic's Claude or GPT-4 remain speculative. The AWS-exclusive launch indicates strategic cloud partnership prioritization over broader distribution.