---
id: 20260617-T0-04
title: "Nemotron 3 Ultra：5500亿参数混合专家模型"
title_en: "Nemotron 3 Ultra: 550B Parameter Hybrid Model"
url: https://ai.daily.yangsir.net/daily/20260617-T0-04
issue_date: 2026-06-17
publish_date: 2026-06-16T04:00:00.000Z
category: research
source_name: "arXiv cs.CL (NLP)"
source_url: https://arxiv.org/abs/2606.15007
---

# Nemotron 3 Ultra：5500亿参数混合专家模型

NVIDIA发布Nemotron 3 Ultra，一个5500亿总参数、550亿激活参数的混合专家模型。该模型基于Mamba-Transformer架构，预训练使用20万亿文本tokens，上下文长度扩展至100万tokens。在智能体推理任务中超越GPT-4，开源权重可供研究使用。开发者可通过vLLM和TensorRT-LLM部署，推理效率提升40%。

## English Version

**Nemotron 3 Ultra: 550B Parameter Hybrid Model**

NVIDIA releases Nemotron 3 Ultra, a 550B total/55B active parameter Mixture-of-Experts hybrid model. Built on Mamba-Transformer architecture, pre-trained on 20 trillion text tokens with 1M token context length. Outperforms GPT-4 in agentic reasoning tasks. Open weights available for research. Deployable via vLLM and TensorRT-LLM with 40% improved inference efficiency.

---

**来源**：[arXiv cs.CL (NLP)](https://arxiv.org/abs/2606.15007)

**详情页**：https://ai.daily.yangsir.net/daily/20260617-T0-04

---

*智语观潮 · Daily — https://ai.daily.yangsir.net/llms.txt*