IFM_AI releases PAN 2, an open foundation world model that predicts physical outcomes before acting

IFM_AI's PAN 2 is an open foundation world model that, per the announcement, predicts physical outcomes of actions rather than text. It uses generative latent prediction with two components: a pre-trained multimodal LLM for conceptual reasoning and a 29B latent diffusion model trained from scratch for perceptual reasoning. Released under Apache 2.0, it includes weights, training and inference code, runnable via vLLM, Hugging Face Diffusers or FastVideo. The model is described as grounded in time, physics and space, supporting multi-step simulation and counterfactual "what if" planning for agents.

✨ AI summary of the linked source — not editorial content.