Cosmos Transfer2.5 2B is a generative video model aimed at producing physics-aware video world states that can be steered through text prompts combined with multiple spatial control inputs derived from real-world captures or simulation outputs. The design centers on a video-to-world workflow where a chosen input source is reshaped into diverse scenarios, allowing developers to customize prompts and synthesize Cosmos video clips suited to downstream physical AI tasks. The supported scenarios highlighted for the model focus on autonomous vehicle simulation and real fleet data as well as robotics, reflecting its intended role in bridging perception research with embodied AI training pipelines that benefit from controllable, scenario-rich video generation rather than purely photorealistic footage.

As a smaller member of the Cosmos Transfer2.5 family, the 2B variant offers a more accessible entry point for teams experimenting with physics-grounded video synthesis, while still accepting text, image, and video conditioning to guide generation. Its open-weights availability makes it straightforward to fine-tune, redeploy, and benchmark locally, which is useful for researchers building custom simulation datasets for autonomy and robotics. The result is a practical tool for prototyping world models that need controllable spatial inputs, particularly in workflows that pair synthetic video with reinforcement learning or imitation learning pipelines for physical AI agents.