NVIDIA Shrank Its World Model to 4B and Dropped the Weights. Cosmos 3 Edge Runs Where Your Games Do.

TL;DR

NVIDIA released the weights for Cosmos 3 Edge, a 4-billion-parameter world model that understands a scene from camera frames, predicts what happens next as video, and generates robot actions, all in one network. The weights landed on Hugging Face on July 20 under the Linux Foundation's OpenMDW-1.1 license, which allows commercial use and derivatives. It is the smallest tier of the Cosmos 3 family, and the only one aimed at Jetson boards and GeForce RTX cards instead of a datacenter.

Frontier-lab world models have so far been things you read papers about. This one is a download.

What actually shipped

The timeline matters, because the announcement and the release were not the same day. NVIDIA introduced Cosmos 3 Edge in a press release from Tokyo on July 15, alongside two new Jetson modules, the T2000 and T3000. The downloadable part, the 4B checkpoint plus a post-trained robot-policy variant, arrived on Hugging Face on July 20, per the release blog and the model card. The press release ran five days ahead of the download link, which by 2026 launch standards counts as restraint.

The family math: 4B is the point

Cosmos 3 launched at the start of June as an omnimodal world-model family spanning text, images, video, and actions. The Cosmos GitHub repo lists three tiers: Super at 64B for H200, B200, and GB200 datacenter hardware, Nano at 16B for RTX Pro 6000 and H100 class cards, and now Edge at 4B. The reference hardware for Edge is Jetson AGX Orin and up; the release blog also lists plain GeForce RTX GPUs and DGX boxes as supported.

At 4B parameters in BF16, the weights are roughly 8 GB before you add context, which is the difference between a robotics lab line item and a card you already own. Your gaming GPU is now, on paper, qualified to supervise a forklift.

Two towers, one world state

Architecturally, Cosmos 3 Edge is a Mixture-of-Transformers: an autoregressive transformer that handles discrete tokens (reading frames, following text instructions, reasoning about the scene) and a diffusion transformer that handles continuous generation (predicting future video), with shared multimodal attention layers between them. Because the towers share attention, the thing that decides what to do and the thing that imagines what happens next are looking at the same internal picture of the world, not passing notes between two separate models.

That shared state is what makes the dynamics tricks work. The model does forward dynamics, given an action, predict its visual consequences, and inverse dynamics, given a desired outcome, recover the actions that produce it. Think of a pool player: forward dynamics is calling where the balls end up after your shot, inverse dynamics is starting from the ball you want sunk and working backwards to the shot. Cosmos 3 Edge runs the table in both directions, and robot-policy training needs both.

The numbers on real hardware

The spec sheet, from the model card and release blog, is refreshingly concrete:

• Real-time control: on Jetson Thor, the model runs at 640x360 control resolution, emits 32 actions per inference, and sustains a 15 Hz control loop.
• Generation: video output at 256p or 480p, 12 to 30 fps, 50 to 150 frames per clip. On an H100 SXM, a 189-frame 480p image-to-video generation takes 27.64 seconds.
• Inputs: text up to 4,096 tokens, images, video, and action trajectories from 16 to 400 steps.
• Embodiments: nine pre-baked action spaces, from 9D camera and vehicle motion up to a 57D egocentric rig, with Franka (10D single, 20D dual arm), UR, WidowX 250, Agibot, and Google robot arms in between.

NVIDIA claims the model ranks first among 4B-class models on its VANTAGE-Bench vision-analytics suite, and that the Cosmos 3 family leads open models on robotics, smart-space, and driving benchmark averages. Both claims are NVIDIA's own; treat them as a vendor's opening bid until third-party evals land.

What is in the box for builders

This is not a weights-only drop. The Cosmos 3 collection includes the base model plus a robot-policy checkpoint post-trained on DROID pick-and-place data, and the GitHub repo carries inference paths for Diffusers, vLLM, and SGLang, along with post-training scripts and distillation recipes. NVIDIA's press release claims developers can adapt the model to a specific robot, vehicle, or sensor rig in about a day of post-training.

The license is the other half of the story. OpenMDW-1.1 is a Linux Foundation license covering the model, weights, and code, permitting commercial use and derivatives with a Built on NVIDIA Cosmos attribution. For a robotics stack, that is materially more open than the research-only or revenue-capped licenses that ate the image-model world.

The Japan subplot

The venue for the announcement was not random. The same press release named FANUC, Yaskawa Electric, Kawasaki Heavy Industries, Sony Group, SoftBank, Fujitsu, Hitachi, NEC, and Kubota among the Japanese industrial names building on Cosmos, alongside robotics startups like GROOVE X, Telexistence, and Enactic. Jensen Huang called physical AI a once-in-a-generation opportunity for Japan. Read together with the 4B edge tier, the strategy is legible: the factory floor does not have an H200, so the model came down to meet the hardware that is actually there.

The caveats, straight-faced

A 640x360 control resolution is enough for manipulation and navigation, not for fine visual detail work. The 27.64-second H100 figure means video prediction is an offline or planning-time tool, not a real-time renderer. Every benchmark number so far is self-reported, the technical report is NVIDIA's own, and VANTAGE-Bench has no independent leaderboard yet. And while OpenMDW-1.1 is genuinely permissive, the attribution requirement means your product is contractually a billboard, small, but load-bearing, for the company that also sells you the GPU it runs on.

Key Takeaways

• NVIDIA released Cosmos 3 Edge, a 4B-parameter world model, with weights on Hugging Face as of July 20 under the permissive OpenMDW-1.1 license.
• It is the smallest of three Cosmos 3 tiers (4B Edge, 16B Nano, 64B Super) and the only one targeting Jetson boards and consumer GeForce RTX cards.
• The Mixture-of-Transformers design pairs an autoregressive reasoning tower with a diffusion prediction tower over shared attention, enabling both forward and inverse dynamics.
• On Jetson Thor it sustains a 15 Hz control loop emitting 32 actions per inference at 640x360; video prediction runs 50 to 150 frames at up to 480p.
• The drop includes a DROID-post-trained policy checkpoint, post-training scripts, and inference support via Diffusers, vLLM, and SGLang.
• Benchmark claims (first at 4B on VANTAGE-Bench, first among open models on robotics and driving averages) are NVIDIA's own and still await independent verification.

Sources: NVIDIA on Hugging Face (release blog), Cosmos3-Edge model card, NVIDIA Cosmos GitHub, NVIDIA press release (July 15), Cosmos 3 research page