ReadingResearchRadarInvestment framework
Sign in / Sign up
← Back to reading
TopicDisconfirmAI EditorialUpdated 2026-10-26 02:42

英伟达Cosmos 3

6 steps2 pieces6 Sources

英伟达于2026年5月31日正式发布并开源了NVIDIA Cosmos 3,这是其首个完全开放的“全模态世界模型”(Omnimodal World Model),旨在解决物理AI中感知、预测与行动分离导致的效率低下问题。【背景】传统机器人开发需串联视觉语言模型(VLM)、视频生成器和策略网络,数据在模型间传递时丢失空间细节且延迟高。Cosmos 3采用创新的**混合Transformer(Mixture-of-Transformers, MoT)**架构,通过“推理塔”(自回归VLM)和“生成塔”(扩散模型)共享潜空间嵌入,在一次前向传播中同时完成场景理解、未来状态模拟和动作轨迹生成。

该系列目前提供两个主要版本:Cosmos 3 Nano(16B参数,面向工作站实时推理)和Cosmos 3 Super(64B参数,面向数据中心高保真仿真),并预告即将推出Cosmos 3 Edge用于边缘设备。英伟达同步成立了“Cosmos联盟”,联合Agile Robots、Runway等公司推进开放世界模型标准。这一举措标志着物理AI从碎片化模型堆叠转向统一基础模型时代,大幅缩短训练周期并降低对真实世界数据的依赖。

How it got here

  1. 2026-05-31

    英伟达开发者博客正式宣布Cosmos 3发布,介绍其双塔MoT架构及Nano/Super模型规格,并开源代码与数据集。

    developer.nvidia.com ↗
  2. 2026-06-01

    英伟达在GTC台北大会上发布新闻稿,宣布成立“Cosmos联盟”,成员包括Agile Robots、Black Forest Labs等,强调其为全球首个完全开放的全模态物理AI基础模型。

    investor.nvidia.com ↗
  3. 2026-06-01

    GitHub仓库发布Cosmos 3正式版标签,确认Nano (16B) 和 Super (64B) 模型权重可用,并列出Text2Image、Image2Video等特定任务变体。

    github.com ↗
  4. 2026-06-03

    第三方媒体WebAInews报道Cosmos 3的技术细节,指出其基于Qwen3-VL初始化权重,并在VANTAGE-Bench、PAI-Bench等基准测试中取得开源SOTA成绩。

Sign in / Sign up
ReadingResearchRadarInvestment framework
webainews.com ↗
  • 2026-10-20

    英伟达机器人生态负责人Amit Goel在播客中解释物理AI架构差异:推理低频而控制高频,Cosmos 3通过双塔共享潜空间解决带宽损失,并指出训练需三台不同计算机分工。

    softwareengineeringdaily.com ↗
  • 2026-10-25

    英伟达Cosmos研究负责人Ming-Yu Liu在节目中阐述设计理念,强调“世界模型”是协同训练的工具集而非单一定义,Cosmos 3将VLM与扩散生成器结合。

    podcasters.spotify.com ↗
  • What is still unsettled

    • Cosmos 3 Edge的具体发布日期及硬件兼容性列表尚未公布。
    • 在极端复杂动态环境(如拥挤街道或柔性物体操作)中,单次前向传播的动作生成准确率是否优于多步规划的传统管线?
    • 开源社区对Cosmos 3微调的实际采用率及衍生模型的质量如何?

    Sources

    • developer.nvidia.com英伟达开发者博客:Cosmos 3技术详解↗
    • investor.nvidia.com英伟达投资者关系:GTC台北发布新闻稿↗
    • github.comGitHub:Cosmos 3发布说明↗
    • softwareengineeringdaily.comSoftware Engineering Daily:Amit Goel访谈↗
    • podcasters.spotify.comMachine Learning Street Talk:Ming-Yu Liu访谈↗
    • arxiv.orgArXiv:Cosmos 3技术报告↗

    Coverage in this topic

    • Ming-Yu Liu, Head of Cosmos Research at NVIDIA: World Models Are a Toolkit, Not a Single Definition2026-10-25News

      Ming-Yu Liu, head of Cosmos research at NVIDIA, explained on the Machine Learning Street Talk podcast how Cosmos 3 combines vision-language models with diffusion generators and proposed that 'world models' are a set of co-trained tools rather than a single definition.

    • NVIDIA Robotics Head Amit Goel: Physical AI Needs Three Computers; Dual-Tower Shared Latent Space Solves the Reasoning-Action Frequency Mismatch2026-10-20News

      Amit Goel, Head of Robotics Ecosystem and Edge AI Product at NVIDIA, explains on the Software Engineering Daily podcast: the fundamental architectural difference between physical AI and LLMs is that reasoning runs at low frequency while motor control runs at millisecond-level high frequency; the two must share context but cannot afford the bandwidth loss of decode-encode. NVIDIA's Cosmos 3 adopts a dual-tower design that exchanges embeddings directly in the latent space; meanwhile, training a world model requires three different computers — for model training, simulation-based reinforcement learning, and edge deployment.