550 billion parameters + millions of tokens context, NVIDIA's open source this time is aimed at long-range agents, cutting costs by 30% is a bit appealing, but the benchmark scores haven't beaten Kimi K2.6 yet, waiting for real-world testing.

NVIDIA open-sourced the flagship large language model Nemotron 3 Ultra, with 550 billion parameters, using Mamba-Transformer mixed MoE, optimized for long-range intelligent agent planning, reasoning, and tool invocation, supporting 1 million token context and achieving 5x throughput with a 30% reduction in inference costs. Third-party benchmark Artificial Analysis scored 48 points, trailing behind Kimi K2.6 with 54 points. Additionally, they released the Nemoclaw orchestration blueprint and the Agent Toolkit for Openshell runtime, now available on Hugging Face, NVIDIA NIM, OpenRouter, and integrated by companies like Glean, serving as an alternative to commercial closed-source large models.