**[NVIDIA Officially Open-Sources Flagship 550B Model Nemotron 3 Ultra: Pioneering Mamba-Transformer Hybrid MoE Architecture, Designed for Long-Range Agents]** According to monitoring by Beating Insights, NVIDIA officially open-sourced its flagship large language model, Nemotron 3 Ultra, on June 4. The model features 550 billion parameters with 55 billion active ones and is optimized for long-range agent tasks such as complex planning, reasoning, and tool utilization. On the third-party benchmark platform Artificial Analysis, Nemotron 3 Ultra scored 48 points on the intelligence index, making it the most powerful open-source weight model in the U.S. to date, second only to Kimi K2.6, which scored 54 points on the Dark Side of the Moon benchmark. From a technical perspective, the model adopts the Mamba-Transformer hybrid expert architecture (MoE), alternating between Mamba-2 state-space model layers and Transformer self-attention layers. This design avoids the quadratic memory bottleneck caused by KV cache growth in ultra-long contexts, enabling support for a 1 million token context window with minimal memory overhead. Compared to dense models of the same scale, the hybrid architecture delivers a 5x throughput improvement on agent tasks while reducing inference costs by 30%. On the ecosystem front, NVIDIA simultaneously released the Agent Toolkit, which includes the NemoClaw orchestration blueprint and the OpenShell runtime. The open-source release includes model weights, datasets, and training recipes. The model is now available on Hugging Face, NVIDIA NIM, and OpenRouter. Enterprise AI search service providers like Glean have already announced integration, using it as an alternative to proprietary commercial large models. [Original Link]