MarkTechPost describes PAN 2 as a 29-billion-parameter model released under the Apache 2.0 license. The report says the system combines a large language model that plans actions with a 29-billion-parameter diffusion transformer.
According to the same report, PAN 2 can simulate up to 20 seconds of action outcomes as video. That capability is presented as part of its role as a world model for physical-AI applications.
The supplied evidence comes from a single MarkTechPost report and does not include an official MBZUAI or IFM confirmation. It also does not establish the physical location or release date; the supplied metadata shows a publication timestamp of November 11, 2026, while the source URL contains November 10, 2026.