NVIDIA has unveiled three new models for working with video — and all of them are available as open-source. We're talking about Cosmos World Foundation Models 2.0, RTX Video, and RTX Diffusion Turbo for Video. All three can be downloaded, run locally on your computer with an RTX graphics card, and used without cloud services.
In short: these are models for generating video from text or images, and each solves its own task. One helps create physically plausible scenes, another improves the quality of existing video, and the third speeds up the generation process. They are all united by one idea — to give developers and enthusiasts tools that work fast, locally, and don't require subscriptions.
Cosmos 2.0: Generation with Physics Awareness 🌍
Cosmos World Foundation Models 2.0 is a suite of models capable of generating video based on text descriptions or images. Their main feature is that they are trained to understand the physical laws of the real world: gravity, lighting, and object movement.
Simply put, if you ask the model to show a ball falling or a car moving, it will try to do so realistically. This is crucial for tasks where you need not just a pretty picture, but a visual simulation — for example, for training robots, testing autopilots, or creating training datasets.
Cosmos 2.0 includes several versions of varying sizes. The most compact one — with 2 billion parameters — can run even on laptops with an RTX 4070. Larger versions, with up to 17 billion parameters, require more powerful hardware but deliver better quality and higher resolution.
The models support various output formats: you can generate short clips in resolutions up to 1280×720, control video duration, and define style via text or a reference image. All weights are available on Hugging Face, and NVIDIA provides code examples for a quick start.
RTX Video: Upscaling Without the Cloud 🎬
RTX Video is a model for improving video quality. It can increase resolution (commonly known as «upscaling»), remove compression artifacts, and sharpen the image.
Unlike cloud services, RTX Video works locally. This means you don't need to upload video somewhere on the internet, wait for processing, and pay for it. Everything happens on your computer if it has an RTX series graphics card.
The model is optimized for RTX Tensor Cores and uses NVIDIA technologies, such as TensorRT, for acceleration. As a result, processing is fast, and you can even work with footage in real-time — for example, enhancing streams or webcam recordings.
RTX Video is also released openly, and it can be integrated into your applications. NVIDIA provides instructions and examples for developers who want to add video enhancement features to their projects.
RTX Diffusion Turbo: Fast Video Generation ⚡
RTX Diffusion Turbo for Video is a model that speeds up the process of generating video from text or images. Usually, diffusion models require many computational steps to get a high-quality result. RTX Diffusion Turbo reduces this number.
While standard models might take 50 steps to create a single frame, Turbo manages it in 4–8 steps. This means generation is several times faster, while quality remains acceptable for many tasks.
The model is especially useful when you need to quickly test an idea, generate a draft, or create many variations in a short time. For instance, for prototyping, creating previews, or testing concepts.
RTX Diffusion Turbo is also optimized for RTX graphics cards and is available for download. It works with the same input data as standard diffusion models but produces results faster.
Local Generation — Why It Matters
All three models share one principle: they run on your computer, without the cloud. This offers several advantages.
First, there's speed. No need to wait for data to be sent to a server, processed, and returned. Everything happens locally, and if you have a powerful enough graphics card, processing is almost instantaneous.
Second, there's privacy. Videos and images do not leave your computer. This is important if you work with sensitive data or simply don't want to transfer content to third parties.
Third, there are no subscriptions. You pay only for the hardware and use the models for free. There are no limits on the number of generations, no queues, and no monthly bills.
Of course, local generation requires a powerful computer. But if you already have an RTX graphics card, you can use it for tasks that were previously available only via the cloud.
Who This Might Be Useful For
These models are aimed at developers, researchers, and enthusiasts. Here are a few examples of where they might come in handy.
Content creators can use Cosmos 2.0 to generate draft videos, RTX Video to improve recording quality, and RTX Diffusion Turbo for rapid idea prototyping.
Game and simulation developers can apply Cosmos to create training data, generate environments, or visualize physical processes.
Researchers in computer vision and machine learning gain access to models that can be fine-tuned, modified, and embedded into their pipelines.
Studios and independent projects working with video can use RTX Video to enhance archival materials or upscale old footage.
Open Source and Availability
NVIDIA has released all three models as open-source. This means you can not only use them as is but also study the code, fine-tune models on your data, adapt them to your tasks, or embed them into commercial products.
Model weights are available on Hugging Face, where you can also find usage examples and documentation. NVIDIA also provides tools for optimizing models for RTX graphics cards — these are the TensorRT and PyTorch libraries that help accelerate performance.
Minimum requirements depend on the model. The lightest versions of Cosmos 2.0 run on laptops with an RTX 4070; larger models will require desktop graphics cards like the RTX 5080 or 5090.
What's Next
The release of these models continues the trend toward local generation and open tools. Until recently, video generation was available only through cloud services with closed models. Now, more and more solutions are appearing that can be launched on your own computer.
This doesn't mean cloud services will disappear. They are still convenient when maximum performance is needed or there's no access to powerful hardware. But local models offer an alternative — for those who want to control the process, experiment with code, or work without the internet.
For now, these models require certain skills to set up and run. But over time, simpler interfaces and ready-made applications will likely appear, making them more accessible to a wider audience.