MBZUAI releases PAN 2 weights and code, with key training stages missing

Mohamed bin Zayed University of Artificial Intelligence’s Institute of Foundation Models released PAN 2, a video ‘world model,’ with model weights, training and inference code, and documentation on November 11. PAN 2 is designed to simulate what happens after a user provides a starting scene and describes actions in text, according to the institute. The release follows other efforts to provide world-model infrastructure, including NVIDIA’s expansion of Cosmos training and deployment tools.

The weights are publicly listed on Hugging Face. The public GitHub repository includes train.py, generate.py, configuration files, launch scripts and documentation. IFM licenses PAN 2’s own code and weights under Apache 2.0, while Wan 2.2, Qwen3.5-9B, Diffusers and other third-party components retain their own licenses. However, the repository says its training code covers pretraining and supervised fine-tuning, but not the reward-learning and model-soup stages used to produce the released weights. That caveat narrows the press release’s description of ‘full training code’: the published materials do not establish end-to-end reproducibility of the released checkpoint.

PAN 2 splits generation between conceptual and perceptual components. An LLM-based conceptual reasoner turns the requested actions into a structured description of events, timing, actors and camera movement. A frozen Qwen3.5-9B text encoder converts that description into conditioning signals for a 29.2 billion-parameter MMDiT diffusion backbone, which predicts video latents. A Wan 2.2 VAE then decodes those latents into video. Put simply, language specifies what should happen and when, while the diffusion model supplies motion, appearance and other visual detail.

The released BF16 diffusion checkpoint supports text-to-video and image-to-video generation at 704 by 1248 pixels and 24 frames per second, with clips up to 20 seconds. The Qwen3.5-9B encoder and Wan 2.2 VAE must be downloaded separately. The repository also documents a packed data loader for precomputed video latents, FSDP2 hybrid sharding, optional context parallelism and Muon optimization. Reading prepared latents from local storage avoids pixel decoding and VAE passes during training.

In MBZUAI’s own multi-step evaluation, PAN 2 completed 75.1% of individual steps and 34.0% of full plans, compared with 75.1% and 32.5% for MiniMax-H3. The institute says the evaluation covered 300 single-action cases and 100 everyday multi-step tasks in image-to-video and text-to-video forms, using automated vision-language-model judges and geometry checks. The company-run results have not been independently reproduced. They therefore do not independently establish MBZUAI’s broader claim that PAN 2 reliably reasons about physical consequences.

The model card also documents substantial limits. It says about two in three tested plans omit at least one step. Appearance can drift between clips in long interactions, text provides only coarse control over dexterous actions, and generated sequences can violate physical constraints—for example, by showing pasta in a colander before it is drained. The project page lists a PAN 2 research paper as ‘coming soon,’ leaving the technical blog, model card and repositories as the current technical record.