MBZUAI's PAN 2 World Model Ships Its 29B Video Backbone and Full Training Code Under Apache 2.0

The Institute of Foundation Models (IFM), the research lab set up by Mohamed bin Zayed University of Artificial Intelligence (MBZUAI), has released PAN 2, an open world model that simulates in video what happens after an action described in words. IFM announced the model from Sunnyvale, California on November 10, 2026, and published the weights on Hugging Face and the code on GitHub under the Apache 2.0 license.

Give PAN 2 a starting scene and an action, and it generates a video of the likely result. IFM calls this “simulative reasoning”: an agent can try out candidate actions inside the model and choose the plan whose simulated outcome looks best, without slow, costly or irreversible trials in the real world.

“A language model learns the world as it has been written down,” said Eric Xing, IFM’s founder and MBZUAI’s president. “That is book knowledge, and it takes you a long way, but you cannot learn to swim from a book.”

A planner that writes, a diffusion model that renders

PAN 2 splits the job between two systems. A large language model acts as the “conceptual reasoning” backbone. It reads the observation and the requested actions, then writes the next world state as a structured, time-stamped description of events, actions tied to specific actors, and camera movement.

The second half is a 29-billion-parameter multimodal diffusion transformer (MMDiT), which IFM calls the “perceptual reasoning” backbone. According to the model card, it was trained from scratch on hundreds of millions of videos annotated in the same structured format. It fills in the detail the text leaves open, such as how a wave breaks or how a hand grips a pipette. A frozen Qwen3.5-9B model encodes the text, and the Wan 2.2 VAE decodes the predicted latents into video.

The design follows IFM’s Generative Latent Prediction (GLP) architecture. The model predicts change in an internal representation and then reconstructs what would be seen, so its predictions can be compared with real footage.

The released checkpoint handles both text-to-video and image-to-video. By default it produces 704×1248 video at 24 frames per second, with clips up to 20 seconds. The sampler defaults to 50 steps with classifier-free guidance of 2.5. PAN 2 is also supported in Hugging Face Diffusers and vLLM.

Any LLM can do the planning

IFM does not tie PAN 2 to one language model. The model card says any instruction-following LLM can serve as the conceptual reasoner, and the repository includes a prompt-rewriter skill. Most studies in IFM’s technical report use Kimi K3. The reasoners it evaluated, in order of full-plan completion, were Claude Opus 5.5, GPT-6 Astra at medium effort, Kimi K3 and Qwen 3.8 Flash.

IFM notes that few world models ship with their full training code. It is publishing the video model weights with full training and inference code, configuration files and documentation. “We want researchers to be able to test PAN 2’s capabilities, understand its limits, and build on it,” said Hector Liu, director of IFM’s Silicon Valley lab.

Where it ranks and where it breaks

IFM’s results are self-reported. On six core capabilities, measured on 300 single-action test cases (physical consistency, causal coherence, entity persistence, constraint satisfaction, viewpoint invariance and objective dynamics), IFM says PAN 2 placed just behind MiniMax-H3 and ahead of four other models. On a set of 100 everyday multi-step tasks, it completed more full plans than any other model tested. On the public Physics-IQ Verified leaderboard, IFM reports that PAN 2 ranks third overall and first among open-weight models. That entry used Claude Opus 5.5, a closed model, as the reasoner.

The model card is unusually direct about failure. “Long plans fail often,” it says: about two in three simulated plans contain at least one step that is not carried out. Appearance drifts across clips in long interactions because each clip starts from at most one frame plus text. Text describes fine-grained hand movements only coarsely. And the model can take shortcuts that break physics, such as pasta appearing in a colander before it has been drained.

A guardrail shipped separately

Because PAN 2 can produce realistic footage, IFM strongly recommends running it behind a separate pan2-guardrail project. That project screens requests with Qwen3Guard-Gen-4B and checks generated video with a Q16 frame classifier. In IFM’s internal tests with Kimi K3 as the reasoner, the guardrail cut the share of harmful text-to-video requests that produced an unsafe video from 33.9% to 1.5%, and no unsafe image-to-video outputs were seen. It also blocked 11.5% of benign text-to-video requests and 3.3% of benign image-to-video requests.

The model card lists research and development in world modeling, agent and robotics simulation, and video for education and creative work as intended uses. It rules out deepfakes of real people and presenting generated clips as real events.