Skip to content

AI

NVIDIA drops four 16B world-sim models, and robots get a dream sandbox

NVIDIA put four new open-weight Cosmos 3 checkpoints on the table, each with 16 billion trainable parameters. They cover driving scenes, sensor streams and robot arms, under a license that allows commercial use.

Cover photo: Köf3 / CC BY-SA 3.0

TL;DR

  1. NVIDIA released four open-weight Cosmos 3 checkpoints for world simulation, each with 16B trainable parameters.
  2. Targets: driving data, camera plus LiDAR streams, simulator-to-video, and bimanual robot action prediction.
  3. OpenMDW1.1 license permits commercial and non-commercial use, under its terms.

What happened

NVIDIA added four open-weight checkpoints to its Cosmos 3 family, part of the Cosmos platform for building world models. All four are post-trained versions of Cosmos 3 Nano, and each one has 16 billion trainable parameters.

The lineup is specialized. Cosmos3-Nano-Transfer-Auto produces multi-view driving data with different weather, lighting and times of day. Cosmos3-Nano-Sim-Auto streams matched camera and LiDAR observations for driving simulation.

Cosmos3-Nano-Sim-Transfer takes controls from a simulator plus a text prompt and outputs photorealistic video. Cosmos3-Nano-Sim-Bimanual predicts video based on robot or camera actions.

The license is OpenMDW1.1, which allows both commercial and non-commercial use under its terms. NVIDIA also points to a technical report and demos alongside the downloads.

Why it matters

A world model is, roughly, a machine that imagines what happens next. For robotics and autonomous driving, that matters because real-world testing is slow, pricey and occasionally involves a fender.

NVIDIA frames the checkpoints as complementary. Sim-Bimanual predicts the visual consequences of actions, while Sim-Transfer converts a physics simulator's controls into generated observations. Two different routes to the same destination: a convincing fake world to train in.

The weather-and-lighting checkpoint is the quiet crowd-pleaser. Rain at midnight is hard to schedule on a test track and trivial to request from a model.

Open weights from a major chip maker also lower the barrier for researchers and startups who would rather not negotiate a proprietary license before teaching an arm to pick things up.

What it means for Europe

Europe has plenty of robotics labs and autonomous driving programs, and open weights with commercial permission are useful to them. Nothing in the release spells out EU-specific terms, pricing or availability, so check the license text against your own compliance needs before building anything on it.

What to do with it

  • Read the license first.OpenMDW1.1 allows commercial use under its terms, so skim those terms before you ship.
  • Match checkpoint to job.Driving teams look at the Auto models, manipulation teams at Sim-Bimanual, simulator users at Sim-Transfer.
  • Start with the technical report.It comes with the release and will tell you more than any headline, including this one.
  • Watch the demos before downloading.Judge the output quality on your own scenario type, not on the showreel.

Ciao, sapiens. The robots now have a dream world. You still have to pay rent in this one.