[AINews] NVIDIA Cosmos 3、Nemotron 3 Ultra 和 RTX Spark
英文摘要
This issue of AI News covers NVIDIA's major open-source announcements including Cosmos 3, a family of omnimodal world models unifying language, image, video, audio, and action; Nemotron 3 Ultra, a 550B open-weight LLM that claims top US open model status; and the RTX Spark personal AI computer. Additionally, MiniMax M3 and Qwen3.7-Plus expand the open agent model field. The news also highlights a shift from model calls to agent runtimes, with products like Perplexity's Search as Code and Google's Managed Agents.
中文摘要
本期AINews报道了NVIDIA的主要开源发布，包括Cosmos 3——一个统一语言、图像、视频、音频和动作的全模态世界模型系列；Nemotron 3 Ultra——一个550B参数的开源LLM，声称是美国最强开源模型；以及RTX Spark个人AI计算机。此外，MiniMax M3和Qwen3.7-Plus扩展了开源智能体模型领域。新闻还强调了从模型调用到智能体运行时的转变，如Perplexity的Search as Code和Google的Managed Agents。
关键要点
NVIDIA launched Cosmos 3, an open family of omnimodal world models using a Mixture-of-Transformers architecture combining autoregressive reasoning and diffusion generation.
NVIDIA发布了Cosmos 3，这是一个开源的全模态世界模型系列，采用混合Transformer架构，结合了自回归推理和扩散生成。
Nemotron 3 Ultra is a 550B open-weight MoE LLM that achieves strong benchmark performance and high inference speed, positioning as the top US open model.
Nemotron 3 Ultra是一个550B参数的开源MoE LLM，在基准测试中表现强劲且推理速度快，成为美国最强开源模型。
RTX Spark is a personal AI supercomputer with up to 128GB unified memory and 1 petaflop FP4 performance, partnering with Microsoft and others.
RTX Spark是一款个人AI超级计算机，拥有高达128GB统一内存和1 petaflop FP4性能，与微软等合作。
MiniMax M3 and Qwen3.7-Plus represent new open-weight models focused on agent capabilities with long context and multimodality.
MiniMax M3和Qwen3.7-Plus代表了新型开源模型，专注于长上下文和多模态的智能体能力。