Roboflow joins NVIDIA’s Cosmos Coalition

Roboflow announced its membership on October 22, 2026. It also says Cosmos 3 post-training is available on its platform, starting with video reasoning. This adapts an already trained model to a particular task or environment: for example, interpreting progress through assembly work. Roboflow connects the collaboration to its Labs research, whose tasks and benchmarks draw on what developers are trying to solve.

The October 20 training guide starts with an Action Recognition project: upload clips, label action time ranges, create a dataset version and select Cosmos 3 Edge. Before training, ensure the sampling window covers the longest action. Completion brings an email and a Models entry. For a new clip, the guide’s Serverless Cloud example requires both the model ID and a Roboflow API key. It returns an action timeline to compare with the expected procedure, checking step presence, order and target duration.

NVIDIA introduced the coalition on May 31. Its members contribute models, research and evaluation methods while using shared Cosmos 3 technologies, training tools and DGX Cloud infrastructure. NVIDIA presents faster innovation and broader interoperability as goals of this collaboration. These are stated aims, not measured results of Roboflow joining.