NVIDIA releases four open Cosmos 3 simulation models
NVIDIA published four new open-weight checkpoints for its Cosmos 3 model family, according to a post on Hugging Face dated November 5, 2026. All four are post-trained versions of Cosmos 3 Nano and carry 16B trainable parameters. What was released Cosmos3-Nano-Transfer-Auto produces synthetic driving data offline. It can generate synchronized video across up to 11 camera views, fisheye cameras…
On November 5, 2026, NVIDIA introduced four open-weight checkpoints for its Cosmos 3 model series, according to a Hugging Face announcement. Each of the four models, post-trained on Cosmos 3 Nano, boasts 16 billion trainable parameters. One model, Cosmos3-Nano-Transfer-Auto, generates synthetic driving data offline, capable of producing synchronized video from up to 11 camera perspectives, including fisheye cameras, with adjustments for weather, lighting, and time of day based on text prompts.
Another model, Cosmos3-Nano-Sim-Auto, serves as the streaming counterpart to the former, outputting brief segments of camera and LiDAR data through denoising steps, enabling prolonged simulations that react to changes in controls. Cosmos3-Nano-Sim-Transfer accepts spatial data from physics simulators, such as depth, segmentation, and edges, along with a text prompt, to generate color video that a policy can utilize.
Meanwhile, Cosmos3-Nano-Sim-Bimanual forecasts the subsequent video frames from robot or camera movements, accommodating two-armed configurations for the YAM and AgiBot models, and can be operated by a policy or joysticks. The released weights adhere to the OpenMDW1.1 license, allowing both commercial and non-commercial uses. However, to operate the Bimanual checkpoint, NVIDIA requires Linux, an NVIDIA GPU, Docker, and the NVIDIA Container Toolkit, noting that the demonstration was tested on a one GB200 system, which is not the minimum memory requirement.
NVIDIA has also made available post-training recipes detailing four stages, ranging from domain adaptation and distillation to training the model on its generated history. Teams can adapt the Transfer-Auto model to their specific camera and LiDAR setup, then distill it into a streaming version. The company cautions that while these models are promising, significant challenges remain in measuring their effectiveness in policy learning and evaluation.
This release targets robotics and autonomous-driving developers but does not impact the sales, support, or operations processes commonly automated in B2B teams.
Written by urgent.news from Dev.to's reporting — not their text. Machine-written — may contain errors; check the original before relying on it.
