IBL News | New York
Nvidia launched Cosmos this week, an open-source video model to generate synthetic data at scale, addressing the data scarcity in robotics.
Trained on 20 million hours of video data and 9,000 trillion tokens, Nvidia Cosmos provides tools for creating physics-aware video simulations for various applications, mainly robotics and autonomous vehicles.
It contains Difusion-based and Autoregressive model architectures, which, respectively, generate continuous tokens to create controllable visual simulations and use discrete tokens to predict future video frames.
It supports text-to-video and video-to-video generation.
Cosmos is available on Hugging Face, licensed under Apache 2, while preview access is offered at build.nvidia.com.
Personal Supercomputer for AI
On the other hand, Nvidia also introduced a personal AI supercomputer, DIGITS, for rapid prototyping of open-source Llama models.
With the New GB10 Superchip, DIGITS debuts as the world’s most miniature AI supercomputer capable of running 200B-parameter models.
It will be available from Nvidia and top partners in May, starting at $3,000.
“By making @AIatMeta Llama models open source, we’re committed to democratizing access to cutting-edge AI technology. With Project DIGITS, developers can harness the power of Llama locally, unlocking new possibilities for innovation and collaboration.”
— @Ahmad_Al_Dahle, VP,… https://t.co/4SAABLaTd4— NVIDIA AI Developer (@NVIDIAAIDev) January 7, 2025
