What happened
The PAN 2 repository describes the project as a general-purpose world model. Given an optional image, text describing the current situation, and text describing actions, PAN 2 generates a video of how the world might change. The README says the system supports simulative reasoning: an agent can generate possible outcomes for candidate actions and choose among them. The repository identifies the Institute of Foundation Models (IFM) at MBZUAI as the developer.
The repository separates PAN 2’s conceptual reasoner from its perceptual reasoning backbone. The former uses a large language model to write a structured description of the next world state, including timed events, actors’ actions, and camera movement. The latter is a 29-billion-parameter multimodal diffusion transformer, or MMDiT, that predicts video latents; a Wan 2.2 VAE decoder turns those latents into video. The README says the public training code covers the perceptual backbone, alongside inference code and the conceptual reasoner’s prompt and agent skill.
For training, the repository describes three stages: 480p pretraining, 720p continued pretraining, and 720p supervised fine-tuning (SFT). Its shipped configuration files cover dense 720p pretraining or continued pretraining and SFT warm-started from a pretraining checkpoint. The README states directly that “Reward learning and the model soup that produce the released weights are described in the technical report and are not part of the training code.” In other words, the public code covers two training stages, while two further parts of the recipe for the released weights are documented in the technical report rather than implemented in this repository.
The README also gives a path for running and adapting the code. Training uses precomputed video latents, CSV metadata, YAML configurations, and Slurm launch scripts. The quickstart describes a one-node debug run of the 29B model, while the documentation covers data preparation, training at scale, checkpoint resumption, and context parallelism. For inference, the launcher runs one worker per GPU and divides prompts among them. Its listed defaults include 50 sampling steps, 704×1248 resolution, 24 frames per second, and support for videos up to 20 seconds. The project’s own code and weights are released under Apache 2.0, according to the README.
What the archive adds
New Runtime’s earlier coverage, “Focusing on Post-Training” (September 28, 2026), argued that post-training is shifting from marginal model polish toward direct control of inference economics. PAN 2 adds a specific release-transparency detail to that broader discussion: some post-training stages used for the released weights are described in a technical report but are not part of the public training code. This is a new fact on a related story, not a restatement.
Trend signal
This does not yet fit a pattern New Runtime tracks. The repository documents a concrete gap between the stages covered by public training code and the stages used to produce released weights, but one model release is not enough to establish a wider trend in how labs publish training recipes. The distinction is still useful when comparing claims about open code, open weights, and reproducibility: those labels can describe different degrees of access to the process behind a checkpoint.
Why now
The repository’s citation identifies PAN 2 as a 2026 project, and its README presents the training code, inference tooling, model information, and technical report as parts of one release. The training-stage distinction is therefore relevant when builders assess what the public repository makes reproducible, rather than treating a repository release as a complete account of how the released weights were produced.
Why it matters
A runnable training pipeline and a complete recipe for producing a released checkpoint are different things. Teams can use PAN 2’s code to inspect or adapt the described pretraining and SFT stages, but the README makes clear that this code does not implement reward learning or model soup. Anyone planning to reproduce the released weights should account for those missing code stages rather than assuming that launching the published scripts is sufficient.
This also sharpens how teams should assess model openness. Weights, inference code, training code, and a description of the full training process each answer different questions. For procurement, research, and internal adoption, the practical question is not simply whether a model is open-weight or whether a repository contains training code; it is which parts of the path from data to released checkpoint are actually executable and which depend on descriptions elsewhere.
Limits and uncertainty
The repository states that reward learning and model soup are absent from its training code and described in the technical report. The README does not, in this account, spell out the exact procedures, configurations, data, or checkpoints used for those stages. The distinction does not by itself show that the released weights cannot be reproduced, nor does it establish how much those stages changed model performance. Those questions would require examining the technical report and independently testing the released code and weights; this note does not verify such a reproduction.
Reader takeaway
Treat the PAN 2 repository as code for pretraining and SFT, not as a complete executable recipe for producing the released weights. Before budgeting for a reproduction or adapting the release, identify how the reward-learning and model-soup stages described in the technical report would fit into your own pipeline—and what evidence would show that your result matches the published checkpoint.