NVIDIA Releases Cosmos3 Super Text2Image Open Weights Model

Introduction

On May 1, 2026, NVIDIA officially entered the competitive text-to-image arena with the release of Cosmos3-Super-Text2Image. While NVIDIA has long dominated the hardware layer of the AI revolution, this release marks a significant strategic move into high-end generative model weights, providing developers with a powerful tool for visual synthesis.

The model is part of the broader Cosmos3 world-model ecosystem, which includes the H-Dreams embodied models showcased at GTC Berlin. Unlike many proprietary competitors, NVIDIA has opted for an open-weights approach, making the model available on Hugging Face for researchers and engineers to fine-tune and deploy on their own infrastructure.

• Release Date: May 1, 2026
• Model Family: Cosmos3 World Models
• Access Model: Open Weights (Hugging Face)
• Primary Focus: High-fidelity text-to-image synthesis

Key Features & Architecture

At the core of Cosmos3-Super-Text2Image is a massive 64B parameter architecture. This scale allows the model to capture intricate semantic nuances and complex spatial relationships that smaller models often struggle with. The architecture is designed to function as a component of a larger 'world model,' meaning its understanding of physics and light is deeply integrated with the Cosmos3 training paradigm.

The model is optimized for high-resolution output and exhibits exceptional prompt adherence. By leveraging NVIDIA's proprietary training techniques used in their physical AI research, the model produces images that are not just visually stunning but structurally coherent, making it a candidate for integration into simulation and robotics training pipelines.

• Parameter Count: 64 Billion
• Architecture: Cosmos3 World-Model integrated
• Optimization: High-fidelity spatial reasoning
• Deployment: Optimized for NVIDIA RTX and enterprise GPU clusters

Performance & Benchmarks

Cosmos3-Super-Text2Image has demonstrated industry-leading performance on the Artificial Analysis Elo scale. The base model achieved an Elo score of 989, signaling its position at the top tier of generative models. Furthermore, the agentic variant, designed for iterative refinement and multi-step prompt processing, pushed this score even higher to 995.

In July 2026, NVIDIA released the Cosmos3-Super-Text2Image-4Step variant, a distilled version optimized for rapid inference. This version achieved an Elo of 973, offering a significant speed-to-quality ratio for real-time applications. These benchmarks were evaluated using the Artificial Analysis methodology, which utilizes human-preference modeling and automated visual quality assessments to rank model capabilities.