NVIDIA Releases Nemotron 3 Super: 1M Token Context, 12B Active Parameters, Built for Agents

NVIDIA has released Nemotron 3 Super, an open-weight model built specifically for agentic workloads. The headline number is the context window: 1 million tokens. At that length, a single inference pass can hold an entire large codebase, a year of business communications, or a multi-document research corpus — the kind of context that multi-step agents need to avoid losing the thread across long task sequences.

The architecture is a hybrid Mixture-of-Experts design. Total parameter count is 120 billion, but only 12 billion are active at any given inference step. That ratio — 10x more capacity than active compute — is what makes large MoE models attractive for enterprise deployment: near-frontier model quality at a fraction of the compute cost. For organizations running thousands of concurrent agent tasks, that efficiency difference compounds into real infrastructure savings.

NVIDIA trained Nemotron 3 Super on 10 trillion tokens of data, published alongside the model weights. That training data release is unusual and operationally important — organizations that need to audit what a model was trained on, for compliance or bias assessment purposes, can do so. Most frontier model providers publish no training data at all.

The model is available on Hugging Face and through NVIDIA’s build.nvidia.com platform. It’s also accessible via Perplexity and OpenRouter for teams that don’t want to self-host. Early enterprise adopters include Cursor for coding assistance, CrowdStrike for security analysis workflows, Palantir for intelligence processing, Oracle Cloud for cloud-integrated agent pipelines, and Zoom for meeting intelligence features.

Nemotron 3 Super sits in the middle of a three-model family. The Nano variant launched in December 2025 for edge and latency-sensitive deployments. The Ultra — the flagship — has not yet shipped. NVIDIA’s positioning throughout has been the Nemotron Coalition, a partner network that treats the model family as infrastructure for enterprise AI rather than a product competing directly with OpenAI or Anthropic.

For developers building multi-step agents, the 1M token context is the practical differentiator. Most agentic frameworks today work around context limits with chunking, summarization, and retrieval layers that introduce latency and error surface. A model that holds the full context natively collapses those architecture layers. Whether Nemotron 3 Super performs well enough at those lengths — without attention degradation in the middle of the context window — will be tested in production over the coming weeks.

The model is released under a permissive open license. Commercial use is allowed without per-deployment royalties, which gives it a meaningful deployment advantage over comparable closed models.