IFM releases PAN 2, an open world model under Apache 2.0

PAN 2 follows the Generative Latent Prediction (GLP) architecture: it predicts an internal state, then the observation that state would produce[1][2].

The conceptual reasoner is a large language model. It writes the next state as a time-stamped description of events, of actions bound to actors, and of camera movement[1][3]. The perceptual reasoner is a latent diffusion model that IFM trained from scratch on hundreds of millions of videos in that same format, chosen for physical dynamics rather than visual appeal, with gameplay, animation and other synthetic footage dropped after the earliest phase[4][1][2].

The backbone is a 29.2 billion parameter MMDiT[3]. A Wan 2.2 autoencoder decodes its latents into frames at 704 by 1248 pixels and 24 frames a second: 241 frames, about 10 seconds, by default, and up to 481 frames, or 20 seconds[3][4]. The text encoder is a frozen Qwen3.5-9B with its last two layers dropped, and MBZUAI has not modified or fine-tuned the Qwen or Wan weights[3][4]. With the diffusion model held fixed, an action sent straight through scored 45.1% on the six-dimension tests and completed no full plan. Full reasoning by Kimi K3 scored 84.6% and completed 34.0% of full plans[2].

Those dimensions are physical consistency, causal coherence, entity persistence, constraint satisfaction, viewpoint invariance and objective dynamism[2]. MBZUAI says that across them PAN 2 placed just behind MiniMax-H3 and ahead of four other models[1]. That comparison covers the open models Cosmos 3 Super, LTX-2.5 and MiniMax-H3, and the closed models FLUX 3 and Seedance 2.5. Closed models were generated once, and blocked or unreturned videos were left out[2].

On 100 everyday tasks, posed as both image-to-video and text-to-video (200 prompts, 456 steps), PAN 2 completed 75.1% of steps and 34.0% of full plans, the highest full-plan rate in the table. MiniMax-H3 matched the step rate and finished 32.5% of full plans. FLUX 3 finished 26.8%, Seedance 2.5 finished 24.5%, Cosmos 3 Super finished 9.5% and LTX-2.5 finished 4.0%[2]. These multi-step figures use Kimi K3 unless IFM says otherwise. The six dimensions use 300 single-action cases[4].

MBZUAI says PAN 2 is third overall on the Physics-IQ Verified leaderboard and first among open-weight models[1]. In the public leaderboard setting, with Claude Opus 5.5 as the conceptual reasoner, the blog says PAN 2 would rank first among open models and third overall. The blog does not print the Physics-IQ score[2].

The video weights, training and inference code, technical blog and documentation are under the Apache 2.0 licence, which covers PAN 2's own code and weights. Wan 2.2, Qwen3.5-9B and Diffusers keep their own licences[1][3][4].

Hugging Face lists the diffusion backbone as 14 safetensors shards totalling 67.37 GB, plus a text encoder of 15.87 GB and a VAE of 2.82 GB. Hugging Face records 86.17 GB for the repository[4]. The install guide also asks for Qwen3.5-9B at about 19 GB and Wan-AI/Wan2.2-TI2V-5B-Diffusers at about 2.7 GB[3]. IFM documents an NVIDIA H200 setup, and an offload mode that streams the diffusion blocks when the bfloat16 model does not fit in free GPU memory. Those guides do not name a minimum GPU memory size. Each inference worker holds the full diffusion model, one worker per GPU[3][4].

The published training code covers pretraining and supervised fine-tuning. Reward learning, and the model soup that produces the released weights, are described in the technical report and are not part of the training code. The blog describes the soup as a weighted average of the checkpoints from before and after reward learning[3][2]. The conceptual reasoner is a language model you run beside the checkpoint. IFM's full-plan order for the ones it tried was Claude Opus 5.5, GPT-6 Astra at medium effort, Kimi K3, then Qwen 3.8 Flash[4].

Long-term world consistency remains an open challenge[1]. Each clip starts from at most one frame plus text, so appearance drifts, and a world that stays consistent through a long exploration still needs memory on both sides of the model[4][2]. About two in three simulated plans miss at least one step, in a study of plans designed to succeed. Text states an action such as opening a drawer more clearly than where each finger meets an object[4].

"A language model learns the world as it has been written down. That is book knowledge, and it takes you a long way, but you cannot learn to swim from a book," said Professor Eric Xing, president of MBZUAI and founder of IFM. "Physical intelligence begins when a system can simulate what its actions would do. PAN 2 is our foundation for that."[1]

Hector Liu, director of IFM's Silicon Valley lab, said researchers should be able to test PAN 2, understand its limits and build on it. "That means releasing the model weights alongside the training and inferencing code."[1]