NVIDIA Officially Opens Download for Cosmos 3 World Model Weights
NVIDIA has officially opened the download for Cosmos 3 world model weights, releasing the first two versions: Super (64.6 billion parameters) and Nano (15.7 billion parameters).
The model is based on a hybrid Transformer architecture and supports full-modal understanding and generation of text, images, videos, sounds, and actions. It is now available on Hugging Face (non-gated) and build.nvidia.com, and supports NVIDIA NIM microservice deployment.
In terms of market mechanisms, physical AI and robotics developers are accelerating the adoption of open world models, with funding concentrating on full-modal foundational model platforms. NVIDIA benefits from attracting developers through ecosystem openness, while closed model suppliers face pressure from open competition.
Source: Public Information
ABAB AI Insight
NVIDIA has previously laid out its strategy for world simulation through the Cosmos series, and the opening of Cosmos 3 weights continues its path from closed-source leadership to strategic openness. The Super version focuses on high-precision physical simulation to support post-training for robotics/autonomous driving, while the Nano version targets low-latency video action inference, aiming to become the infrastructure for physical AI.
In terms of capital strategy, NVIDIA is fully opening model weights for free download and post-training, motivated by expanding NIM deployment and CUDA usage through proprietary model building by developers, creating a positive feedback ecosystem for physical world simulation rather than short-term model sales.
Similar cases include the rapid ecosystem expansion following the opening of the Llama series and NVIDIA's early CUDA open strategy; currently, physical AI is in the early stages of transitioning from closed simulation to open full-modal world models.
Essentially, this represents a technological shift: world models are moving from closed, high-cost tools to fully open full-modal foundational models. The mechanism is driven by the explosive demand for high-quality synthetic data in physical AI, allowing open weights to quickly lower R&D barriers, thereby enabling NVIDIA to gain long-term infrastructure pricing power through ecosystem control.
ABAB News · Cognitive Law
Open weights are not about profit-sharing but are the starting point for building a new ecosystem.
The physical world needs open models; closed models struggle to become infrastructure.
Excellent companies sell platforms, while closed companies sell individual models.