要点速览
- NVIDIA launches Nemotron 3 Ultra and Cosmos 3 multi-modal world models, with industry-leading performance and open-source access
- Adobe Photoshop/Premiere get 2x speed boosts via NVIDIA RTX Spark-powered AI-native workflows
- OpenAI expands AWS availability with Codex and upcoming Daybreak security features
- Perplexity introduces Search as Code architecture for AI agents, improving search efficiency
- Hugging Face hits 1,000 public AI repositories, with NVIDIA leading open-source AI contributions
一、NVIDIA AI Product & Infrastructure Updates
- NVIDIA has fully launched the Vera Rubin multi-rack pod system designed for large-scale agent AI infrastructure, with partners including Microsoft, Dell, and CoreWeave deploying NVL72 engineering racks based on the platform. — via 1
- The company released Nemotron 3 Ultra, the largest model in the Nemotron 3 series with ~550B total parameters and 90% sparsity (55B active parameters). It ranks as the most capable US open-weight model, scoring 48 on the Artificial Analysis Intelligence Index, processing over 300 tokens per second on DeepInfra endpoints, and will offer an NVFP4 quantized version for improved inference performance. — via 1
- NVIDIA Cosmos 3 multi-modal world models are now available, with four variants including 16B Nano and 64B Super. The Super version tops both text-to-image and image-to-video open-weight model rankings on Artificial Analysis, uses a hybrid Transformer architecture with mixed inference and diffusion towers, supports structured JSON prompts, and is fully licensed under OpenMDW 1.1. Weights, code, datasets, and fine-tuning tutorials are live on Hugging Face, with official and third-party APIs and pricing coming in the next few weeks. — via 1 2
- NVIDIA announced RTX Spark-powered updates for Adobe Photoshop and Premiere Pro, delivering up to 2x speed improvements in AI editing, color grading, and effects workflows with full GPU acceleration and AI-native processes, set to launch soon. — via 1
- The company launched the NVIDIA Factory Operations Blueprint (FOX) factory management agent reference design, with early users including Hon Hai, Pegatron, Advantech, and Wistron reporting productivity, quality, and efficiency gains. — via 1
- NVIDIA shared its AI Media Solution, bringing real-time AI performance to live and on-demand media workflows, including a 92% accurate, 22ms low-latency AI video authenticity detector, and RTX Video Super Resolution/frame generation that upscales 480p content to 8K and boosts frame rates by 2-4x, deployable on-premises or in the cloud. — via 1
二、Industry Partnerships & Ecosystem Updates
- OpenAI has made its frontier models and Codex fully available on Amazon Bedrock, allowing enterprises to build products with OpenAI's security, compliance, and governance workflows. The launch marks the start of OpenAI's AWS expansion, with upcoming Daybreak cybersecurity features planned. — via 1
- Hugging Face has surpassed 1,000 public AI repositories, including 820 models, 249 datasets, and 57 spaces, with nearly 60,000 followers. NVIDIA's LocateAnything and PiD rank first and fifth on the Hugging Face trending models list, and the platform has adopted the Linux Foundation's OpenMDW framework. — via 1 2
- Corning's fiber manufacturing engineering team used Cosmos 3's synthetic data generation capabilities, training a defect detection model with 8 real defect images plus synthetic examples to achieve 0.95 average precision and perfect recall on the hardest defect categories, outperforming a baseline trained only on real data. — via 1
三、AI Agent & Tooling Innovations
- Perplexity AI launched Search as Code, a new search architecture for AI agents that generates Python code to directly call search stacks instead of iterative function calls. The system is now available via the Perplexity Agent API and is the default feature for Computer, enabling parallel searches, custom filtering, deduplication, and faster processing with fewer back-and-forth steps. — via 1 2 3
- OpenAI's Codex now includes the Sites feature, which converts user work, ideas, and plans into interactive websites or applications shareable via URL. The feature is rolling out to commercial and enterprise plans first, with broader availability planned later. — via 1
四、Industry Observations & Trends
- A study found that AI coding tools boost output significantly: GitHub Copilot increases code production by 2.2x, early Claude Code local agents by 7.4x, and remote coding agents by 17.3x, though real-world code releases only see a 30% improvement due to human coding bottlenecks. — via 1
- swyx shared insights from the Latentspace Podcast, noting that video model intelligence derives more from language than video data, that the next major AI breakthrough will be video agents rather than better video models, and that generative UI will replace HTML/CSS by translating user intent directly into pixels. — via 1
