IFM 推出开放基础世界模型 PAN 2，接受描述的动作并生成最多 20 秒展示物理后果的视频，定位为让 Agent 把试错成本移入模拟。架构上由语言模型先规划情节、再由 29B 视频模型渲染。官方基准显示 PAN 2 完成单步 71.5%、完整计划 34.0%，在开放模型中完整计划完成率第一、Physics-IQ Verified 排名第三。
Agents learn from trial and error. The expensive part is the error.
A world model moves the error into simulation, which is the whole case for something like PAN 2.
This is a free, open "action-conditioned world model".
Takes an action you describe, and generates up to 20 seconds of video showing what would physically happen next. Its new trick is splitting the job in 2: a language model first works out the plot, then a 29B video model renders it.
在 X 查看被引用的帖子Introducing PAN 2: an open foundation world model for physical intelligence.
A language model predicts the next word. PAN 2 predicts the next world: what happens after each action, grounded in time, physics and space.
Towards AI that simulates “what if?” before it acts.
More details: ifm.ai/pan/
来源：Rohan Paul · x.com