AIAny
AI Model2026
Icon for item

PAN 2

Simulates how a scene changes in response to text-described actions, optionally conditioned on a starting image, and generates videos up to 20 seconds long. It pairs language-model reasoning about events with a 29B diffusion model for visual prediction.

Introduction

A video generator usually tries to depict a requested scene; this model instead treats video as a testable prediction of what an action will cause. That shift makes it interesting for simulative reasoning: an agent can compare imagined outcomes before acting, rather than relying on a plausible-looking clip as evidence of physical correctness.

Key Capabilities
  • Its conceptual reasoner turns observations and actions into a structured, time-grounded description of events, actors, and camera motion. This gives the visual model an explicit account of what should happen and when.
  • A 29B-parameter diffusion backbone predicts video latents from that description, with an optional starting image; output targets 704 × 1248 at 24 fps and can span up to 20 seconds. The split is useful because event planning and visual rendering can be examined separately.
  • The released checkpoint is the perceptual backbone, not a fully bundled end-to-end reasoner. Users supply a structured world-state caption or run a separate language model to rewrite free-text prompts, so the conceptual reasoning quality depends on that component.
Who It Suits—and the Trade-Off

A good fit for researchers studying world models, action-conditioned video, and agent planning, or for testing candidate actions in a visual simulator. Look elsewhere for dependable long-horizon execution or safety-critical control: many multi-step plans miss at least one step, prompts are sensitive to wording, and clips generated from a single starting frame can drift in appearance. Its videos are predictions, not evidence of real events.

Information

  • Websitehuggingface.co
  • OrganizationsInstitute of Foundation Models, Mohamed bin Zayed University of Artificial Intelligence
  • Published date2026/10/07

More Items

Provides six open language models from 0.9B to 375B-A23B for reasoning and agentic tasks. The paper also introduces sparse attention and a decoding speedup that needs no separate draft model.

Hugging Face

Provides first-person video of skilled manual work with synchronized motion-sensor data for robot learning and dexterous manipulation research. The first release spans 10,124 hours from 12,227 contributors, with task-level annotations.

Hugging Face
AI Audio2026

Turns English text into speech with two American English voices, running on CPUs or in a browser. Its 4.5M-parameter model was trained from scratch on synthetic speech.