Nvidia Unleashes Giant Open-Weight Model Amidst Industry Policy Shift

Nvidia's Nemotron 3 Ultra is a massive open-weight language model that has sparked significant interest in the AI community. Released on June 4, 2026, this 550-billion-parameter model is available for download on Hugging Face with no API key required. The hybrid Mamba-Attention Mixture-of-Experts (MoE) architecture used in Nemotron 3 Ultra allows it to process sequences with a fixed-size memory state instead of full attention.

Nemotron 3 Ultra is designed to excel at agentic workloads, including long-running tasks, multi-step tool use, and coding agents that need to hold large amounts of context. The model's technical report explains that the hybrid architecture combines Mamba-2 state-space layers, traditional attention layers, and MoE layers with LatentMoE routing. This results in a more efficient and cost-effective model for processing long contexts.

Benchmark scores show that Nemotron 3 Ultra performs well on agentic and coding tasks but falls behind closed frontier models like Claude Opus 4.8 on general-reasoning tests. The Artificial Analysis Intelligence Index, which measures composite reasoning, coding, knowledge, and agentic tasks, puts Nemotron 3 Ultra at 47.7-48.2, roughly 13 points behind the leading closed models.