Cosmos 3
A family of interactive simulation models from NVIDIA covering multi-view driving, continuous lidar-based driving simulation, controlled scene generation, and dual-arm robot action-to-video generation, among other directions. Checkpoints and post-training recipes are released openly.
Timeline (last 90 days)
The 4 new Cosmos 3 open-weight models from NVIDIA serve different purposes: some can generate multi-view autonomous driving data under different weather, lighting, and times of day; some can output camera imagery and lidar observations simultaneously; some render control signals from a simulator into near-realistic video; and others predict the next frames based on the motions of a dual-arm robot.
NVIDIA published a blog on Hugging Face adding 4 open-weight models to the Cosmos 3 world model family, each with about 16 billion trainable parameters, all obtained by post-training from Cosmos 3 Nano.
NVIDIA's new Cosmos 3 open-weight models use an open license that permits commercial use, and alongside the public weights the post-training scheme was also released.
Nvidia's Cosmos 3 models are responsible for converting control information into imagery, while the physics simulator handles physical laws and ground-truth generation; the two are collaborative rather than substitutive.
Nvidia introduced four Cosmos 3 interactive simulation models on November 5 and released checkpoints and post-training recipes.
Cosmos 3 is available in three sizes—Super, Nano, and Edge—where Super pursues the highest fidelity and Edge is optimized for edge devices such as Jetson Thor, Orin, and DGX Spark.
Daisy demonstrated running the Cosmos 3 video generation model on two DGX Sparks connected via Ethernet, using context parallelism with low-latency NCCL for data communication.
All Facts (7)
The 4 new Cosmos 3 open-weight models from NVIDIA serve different purposes: some can generate multi-view autonomous driving data under different weather, lighting, and times of day; some can output camera imagery and lidar observations simultaneously; some render control signals from a simulator into near-realistic video; and others predict the next frames based on the motions of a dual-arm robot.
50%UnverifiedNVIDIA's new Cosmos 3 open-weight models use an open license that permits commercial use, and alongside the public weights the post-training scheme was also released.
50%UnverifiedNvidia's Cosmos 3 models are responsible for converting control information into imagery, while the physics simulator handles physical laws and ground-truth generation; the two are collaborative rather than substitutive.
50%UnverifiedNVIDIA published a blog on Hugging Face adding 4 open-weight models to the Cosmos 3 world model family, each with about 16 billion trainable parameters, all obtained by post-training from Cosmos 3 Nano.
50%UnverifiedNvidia introduced four Cosmos 3 interactive simulation models on November 5 and released checkpoints and post-training recipes.
50%UnverifiedCosmos 3 is available in three sizes—Super, Nano, and Edge—where Super pursues the highest fidelity and Edge is optimized for edge devices such as Jetson Thor, Orin, and DGX Spark.
50%UnverifiedDaisy demonstrated running the Cosmos 3 video generation model on two DGX Sparks connected via Ethernet, using context parallelism with low-latency NCCL for data communication.
50%