What happened
The Institute of Foundation Models (IFM), established by Mohamed bin Zayed University of Artificial Intelligence (MBZUAI), announced PAN 2 on November 10, 2026. IFM describes it as a general-purpose world model: a user supplies a starting scene and an action in words, and the system simulates a video of the likely outcome, grounded, in the institute’s account, in physics, time, and space. IFM frames this as “simulative reasoning,” allowing people and AI agents to examine possible outcomes before taking an action in the real world.
The release describes two components. For conceptual reasoning, a large language model works out what will happen and when, then produces a time-stamped description of events, actions, and camera movement. For perceptual reasoning, a video diffusion model turns that description into a video and supplies visual details. IFM says it trained this video model from scratch on real-world videos selected for physical dynamics rather than visual appeal. The architecture builds on Generative Latent Prediction, which predicts change in an internal representation and reconstructs what would be observed so that predictions can be checked against the real world.
IFM reports results across several evaluations. On six dimensions—physical consistency, causal coherence, entity persistence, constraint satisfaction, viewpoint invariance, and objective dynamics—PAN 2 placed just behind MiniMax-H3 and ahead of four other models. The announcement says PAN 2 completed more full plans than any other model tested in a multi-step planning evaluation. On the public Physics-IQ Verified leaderboard, it placed third overall and first among open-weight models. These are claims in IFM’s announcement; the release does not provide the underlying scores or evaluation protocol in the text presented here.
The institute says the model can simulate situations absent from its training data, giving an egg dropped on different surfaces as an example, and can be used as a world for agents to explore, such as a kitchen in which an agent practices cooking pasta. IFM also explicitly identifies long-term world consistency as an open challenge, with further detail in its technical blog. The release makes video model weights, full training and inference code, documentation, and a technical blog available under the Apache 2.0 license. IFM says the release is available through its website and can be adapted for fields including robotics and education. It does not state a price in the announcement.
What the archive adds
New Runtime's earlier coverage has followed open world-action models in robotics, including NVIDIA's GR00T 2 and Isaac Lab 3.0 (November 9, 2026) and Verne's V0.5 industrial world-action model (November 5, 2026). PAN 2 is a new release in that broader thread, not a restatement: IFM presents a general-purpose model for simulating action outcomes as video, rather than a robotics-specific control stack.
Trend signal
This release adds to the observable shift toward open models that represent and predict physical change, but it does not yet establish a new pattern New Runtime tracks. IFM's announcement describes a research system and benchmarks, not a production interface or evidence of reliable deployment in applications. In particular, the release's stated long-term consistency challenge cautions against treating generated video as a dependable substitute for testing actions in the world.
Connections to other events
New Runtime's earlier coverage of NVIDIA's GR00T 2 and Verne's V0.5 also concerned models aimed at action and physical environments. PAN 2 meets those releases at the question of how systems can represent what an action changes, but its stated mechanism is simulation through video generation. The announcement does not establish that PAN 2 is integrated with either robotics stack or that its simulated outcomes can be used directly as robot commands.
Why it matters
A model that can produce plausible action outcomes gives a team a potential way to explore scenarios before spending time, hardware, or materials on real trials. The open release is particularly consequential for researchers and robotics builders who need to inspect training and inference code, reproduce experiments, or adapt the model rather than rely on a hosted black box. Apache 2.0 also gives teams a clear license framework for experimentation and modification.
The practical value, however, depends on whether a simulated outcome is sufficiently faithful for the decision being made. Visual plausibility alone is not evidence that a model captures relevant physical constraints or remains consistent across a long sequence. IFM's benchmark claims suggest promising performance on selected tests, but they do not show safety or reliability in a deployed robot. Teams should treat PAN 2 as a candidate simulation and research tool, then compare its predictions with measured outcomes before allowing an agent to act on them.
Limits and uncertainty
The announcement reports rankings but does not give the scores, full evaluation methods, or independent replication of the results in its release text. It does not quantify how often simulated outcomes match real-world outcomes, or how performance varies across environments and action types. IFM itself says long-term world consistency remains an open challenge. The release also does not state hardware requirements, throughput, or a price for use through its website. Availability of weights and code makes inspection possible, but does not by itself establish that the model is efficient or dependable enough for a particular operational setting.
Skill to build now
If your team evaluates physical-action agents, build a small replay set of scenes with actions whose outcomes you can verify. Run a simulator such as PAN 2 on those cases, record predicted outcomes, and compare them with observed results across both single actions and longer sequences. Separate visual similarity from decision-relevant correctness: whether an object remains in place, whether a constraint is violated, and whether the final state supports the intended plan. That gives you a concrete basis for deciding where simulation can safely narrow experiments and where real-world validation remains necessary.
Reader takeaway
PAN 2 is an open research release for generating video simulations of described actions, with training and inference code available under Apache 2.0. Its reported benchmarks make it worth evaluating, not trusting by default. For builders, the important next step is to measure whether its predictions hold up on the physical tasks and time horizons that matter to their own systems.