Nvidia unveiled four major AI innovations including the Neotron 3 Ultra open-source AI model, the Vera CPU optimized for AI agents, the multimodal Cosmos 3 foundation model for robotics, and the RTX Spark super chip combining GPU and CPU for powerful local AI processing. These advancements emphasize openness, performance, and integration to accelerate AI development across various domains, from AI models and robotics to personal computing.
Nvidia recently unveiled four major AI-related innovations at GTC Taipei, starting with the release of Neotron 3 Ultra, their latest open-source AI model. This model boasts an impressive 550 billion parameters and utilizes a hybrid Mamba Transformer mixture of experts architecture, making it five times faster and 30% more cost-efficient than existing open models. Nvidia emphasizes openness by providing not only the model but also the training scripts and datasets, encouraging developers to customize and improve upon it.
Next, Nvidia introduced Vera, a groundbreaking CPU designed specifically for the era of AI agents rather than traditional human-centric computing. Vera features up to 3.6 terabytes per second of internal bandwidth and incorporates the Olympus core optimized for modern data center workloads, including Python runtimes and sandbox code execution. It offers significantly lower memory latency and faster core-to-core communication than conventional CPUs, enabling enhanced GPU utilization and improved AI processing performance.
For robotics and physical AI applications, Nvidia launched Cosmos 3, an open-world foundation model trained on an enormous multimodal dataset including images, videos, sound, text, and action data. Cosmos 3 integrates multiple capabilities such as future prediction, domain transfer, physical reasoning, and action generation into a single versatile model. Available in both a high-accuracy super model and a smaller nano model, Cosmos 3 is open source with accessible weights, training scripts, and datasets to accelerate development in physical AI.
The final and most ambitious announcement is RTX Spark, a collaborative effort between Nvidia and Microsoft to reinvent the personal computer for the AI age. RTX Spark is a single super chip combining a Blackwell RTX GPU with 6,144 CUDA cores and a custom 20-core Grace CPU, delivering one petaflop of AI performance and 128 GB of unified memory. Built on a cutting-edge 3nm process with 70 billion transistors, this platform is designed to run AI agents locally and securely, marking a significant evolution in PC architecture.
Together, these innovations showcase Nvidia’s commitment to advancing AI technology across multiple domains—from open AI models and specialized CPUs to robotics and personal computing. By focusing on openness, performance, and integration with AI agents, Nvidia aims to empower developers and users to harness AI more effectively, driving the next wave of AI-powered applications and devices.