NVIDIA has released Cosmos 3 Edge, a 4-billion-parameter open world model designed to run visual reasoning and action-generation workloads closer to robots and sensors. The model is available through NVIDIA’s Cosmos 3 collection on Hugging Face, together with a manipulation-policy checkpoint and post-training scripts.
The release targets physical AI systems that must understand a changing scene, predict what may happen and select an action without sending every step to a data centre. NVIDIA says Cosmos 3 Edge can operate as a compact vision-language model or as a world action model that connects observations with robot controls.
One representation for perception and action
Cosmos 3 uses two transformer towers. An autoregressive tower processes visual and text tokens for understanding and reasoning, while a diffusion tower handles vision, audio and action tokens for prediction and generation. The towers keep separate normalisation and feed-forward layers but share multimodal attention, which aligns information across the different inputs and outputs.
The model represents actions as compact geometric vectors covering translation, rotation and manipulation state. This common format is intended to work across different physical systems, such as a vehicle, camera, robot arm or gripper. By linking control signals to changes in a scene, the model can predict the visual consequence of an action or infer the action associated with an observed change.
Real-time edge performance claims
NVIDIA says the post-trained world action configuration works with 640 by 360-pixel observations and produces 32 actions per inference on Jetson Thor, reaching 15Hz control. It also claims the model ranks first among similarly sized models on VANTAGE-Bench for vision analytics and delivers state-of-the-art robot-policy learning results.
Those figures are vendor-reported and workload-specific. Developers will need to test latency, accuracy, thermal limits and failure behaviour on their own hardware and in the environments where a robot will operate. Benchmark leadership does not on its own establish safe real-world performance.
NVIDIA lists support across RTX Pro and GeForce RTX GPUs, DGX systems and Jetson devices, including the recently announced T2000 and T3000 modules. The compact size is intended to make deployment possible on memory-constrained systems while retaining enough capability for scene understanding and action reasoning.
Policy checkpoint and training tools
The release includes Cosmos 3 Edge Policy for the DROID dataset, aimed at pick-and-place manipulation, along with scripts for adapting the model. NVIDIA says developers can fine-tune target workloads on a small H100 cluster or DGX Station before deploying the resulting policy to edge hardware.
NVIDIA is also providing post-trained checkpoints and recipes for the broader Cosmos 3 family. A separate Cosmos 3 Super four-step distillation reference reduces diffusion from 35–50 denoising steps to four, with the company claiming up to 25 times faster inference while preserving image and video quality. That 64-billion-parameter reference is distinct from the 4-billion-parameter Edge model but demonstrates the same emphasis on practical optimisation.
Availability and limitations
The model files, collection and supporting scripts can be downloaded from Hugging Face. Open availability gives robotics teams a way to inspect and adapt the model, but the announcement does not remove the need for task-specific validation, safety constraints and fallback behaviour.
NVIDIA says future work will include interactive world generation, driving simulation, robotics-policy improvements and optimisation for frameworks such as vLLM. These are roadmap statements rather than currently available capabilities.
What teams need beyond the model
A deployable robot policy needs more than model weights. Developers must define sensor calibration, control frequency, permissible actions, emergency stops and the conditions that trigger a human takeover. Training data should reflect the lighting, objects, viewpoints and physical variability of the target environment, while evaluation should include rare but consequential failures.
The open scripts make adaptation more accessible, but post-training can also narrow performance or introduce behaviour that was not present in the base checkpoint. Teams should retain a fixed evaluation set, compare the adapted model with the original and record which checkpoint, data and configuration produced each result.
Licensing and model-card conditions should be reviewed alongside technical performance. The announcement describes Cosmos 3 Edge as open, but organisations still need to confirm the applicable licence, acceptable-use conditions and obligations for any datasets or downstream checkpoints they distribute.
On-device operation can reduce network latency and keep sensor data local, but it also shifts responsibility to the edge system. Hardware monitoring, secure updates and recovery after a failed inference become part of the product design. Developers should measure end-to-end control latency rather than relying only on the model’s reported inference rate, since sensor capture and actuator response also affect safety.