Nvidia officially open-sourced its flagship 550-billion-parameter large language model on June 4th — the Nemotron 3 Ultra. Designed for complex planning, reasoning, and tool use in long-horizon agent tasks, it activates just 55 billion parameters per inference.
https://twitter.com/nvidia/status/2062522316672667770
On the Artificial Analysis intelligence benchmark, the model scored 47.7, making it the strongest open-weight model from the US. But it still trails several Chinese open-source models: Kimi K2.6 (53.9), MiMo-V2.5-Pro (53.8), and DeepSeek V4 Pro (51.5).
Architecturally, Nemotron 3 Ultra uses a novel Mamba-Transformer hybrid mixture-of-experts (MoE). By alternating Mamba-2 state-space model layers with Transformer self-attention layers, it sidesteps the quadratic KV-cache memory problem in long contexts, supporting a 1-million-token context window at minimal memory overhead. In agent tasks, the hybrid architecture delivers up to 5x higher throughput and 30% lower inference cost.
On the ecosystem side, Nvidia also released an Agent Toolkit, including the NemoClaw orchestration blueprint and OpenShell runtime. The open-source release includes model weights, datasets, and training recipes. The model is available on Hugging Face, NVIDIA NIM, and OpenRouter. Enterprise AI search provider Glean has already announced it will use the model as a commercial closed-source alternative.