IFM releases open PAN 2 model for physics and robotics under Apache 2.0

The AI lab behind the @IFM_AI account released PAN 2 on 11 November 2026 and calls it “the best open model in physics and robotics”. The lab says PAN 2 ranks first among open models on the Physics-IQ Verified benchmark and that its weights, training code and inference code are all available under the Apache 2.0 licence.

The announcement is short and gives no scores. Everything known about PAN 2 so far comes from that one post, so the model’s claimed rankings should be read as the maker’s own claims until outside testers report results.

What IFM announced with PAN 2

The @IFM_AI post makes three claims:

• Physics-IQ Verified: PAN 2 is the “#1 open model” on this benchmark.
• RBench: PAN 2 is listed as “#1 open models in our RBench evaluation”, a test IFM describes as its own.
• Release terms: “Open weights, training code and inference code under Apache 2.0”.

The post frames PAN 2 around two areas, physics and robotics. It does not describe the model’s architecture, size, inputs, outputs or how long a sequence it can produce. It also does not name the earlier PAN model or say what changed between versions. For a builder, that means the post settles the licence and what files are released, but not what the model actually does or what hardware it needs.

Detail	What the announcement says
Model name	PAN 2
Release date	11 November 2026
Licence	Apache 2.0
Released files	Open weights, training code, inference code
Physics-IQ Verified	#1 open model (no score given)
RBench (IFM’s own test)	#1 among open models (no score given)
Parameter count, inputs, outputs	Not stated

Weights, training code and inference code under Apache 2.0

The most concrete part of the release is what ships. IFM lists three pieces: the trained weights, the code used to train the model, and the code used to run it. All three are covered by Apache 2.0, according to the post.

Each piece serves a different use:

• Weights are the trained parameters. With them, a team can run the model on its own hardware instead of calling a hosted service.
• Inference code is the software that loads the weights and produces outputs. It sets the reference for how the model is meant to be run.
• Training code shows how the model was built. It lets researchers reproduce or extend training, for example by fine-tuning on their own robot or physics data, which weights alone do not make easy.

The post names the licence but does not add further usage terms or conditions. Teams planning commercial use should read the licence files that come with the release before building on it.

Physics-IQ Verified and RBench: what the rankings say

Both ranking claims are limited to open models. The post does not say how PAN 2 compares with closed commercial systems on either test, and it gives no numbers.

Physics-IQ

Physics-IQ is a Google DeepMind benchmark built to check whether video generators have learned real physics. It covers fluid dynamics, optics, solid mechanics, magnetism and thermodynamics. Its authors summed up their findings this way:

“visual realism does not imply physical understanding.”

That conclusion is reported by the newsletter Artificial Impact, which places Physics-IQ alongside other studies that tested photorealistic video generators such as Sora, Runway, Pika, Lumiere and Stable Video Diffusion. IFM cites a “Verified” version of the benchmark. None of the sources explains how Physics-IQ Verified differs from the original Physics-IQ, who maintains it, or how many open models appear on it.

RBench

IFM calls RBench “our RBench evaluation”. It is the lab’s own test, so the ranking there is a self-reported result, not a third-party one. The sources do not describe what RBench measures.

Why physics tests matter for video and robotics models

IFM’s choice to lead with physics rankings fits a wider debate about whether generative models understand how objects behave or only how they look. Artificial Impact reports several results that make the point:

• ByteDance’s phyworld study found that video generation “fails to learn physical laws from video data, even with scaling,” with errors jumping by an order of magnitude on scenarios unlike the training data.
• Meta’s IntPhys 2 benchmark shows short videos where an object does something impossible and asks whether the scene is possible. Humans answer correctly 85 to 95% of the time, while current models score at or near chance.

The same newsletter notes that roughly $4 billion flowed into world-model companies in about a year, with labs such as AMI Labs and World Labs and robotics firms including Skild AI and Physical Intelligence raising large rounds. It also notes that these companies give different answers when asked what a world model is. IFM’s post does not use that term for PAN 2; it describes the model as an open model for physics and robotics.

Researchers outside video generation are working on the same gap. At MIT’s Computer Science and Artificial Intelligence Laboratory, a team with Tsinghua University built GeoPT, a pretraining approach that teaches simulation models physics through synthetic particle interactions with 3D shapes. Co-lead author Minghao Guo put the motivation plainly:

“We believe physics is the third modality for AI models, after text and pixels.”

That quote comes from Tech Xplore’s coverage of the MIT work. GeoPT is a different kind of system from PAN 2, but both reflect the same push to get physical behaviour right, not only appearance.

For related coverage of physical AI, see Unicorn API’s report on DeepMind’s Gemini Robotics 2.5 and its open robot safety benchmark, and for open generative video, Lightricks’ LTX-2.5.

What PAN 2 means for builders

For developers working on robotics, simulation or physically grounded video, the release terms are the clearest takeaway. Open weights plus training and inference code under Apache 2.0 means a team can, in principle, run PAN 2 locally, inspect how it was trained and adapt it to its own data, subject to the licence terms that ship with it.

The claims to verify before relying on the model are these:

1. The rankings are self-reported. Both the Physics-IQ Verified and RBench claims come from IFM, and RBench is IFM’s own evaluation. No independent results for PAN 2 are in the sources available for this article.
2. “Best open model” is a narrow comparison. The post compares PAN 2 only with other open models. Teams weighing it against hosted closed systems will need their own tests.
3. Physics benchmarks are a high bar. The research summarised by Artificial Impact suggests strong visuals and physical accuracy often diverge, and that models struggle most on situations unlike their training data. A top open-model rank does not by itself show how PAN 2 handles a specific robot, scene or task.
4. Practical specs are missing. Without a stated size, input format or output length, teams cannot yet estimate hardware costs or latency from the announcement alone.

Open questions and what to watch next

The announcement leaves several points open. The sources do not say:

• how large PAN 2 is or what hardware it needs to run;
• what it takes as input and what it produces;
• what scores it achieved on Physics-IQ Verified or RBench, and which models it was ranked against;
• how Physics-IQ Verified differs from the original Physics-IQ benchmark described by Google DeepMind;
• what RBench measures and whether it will be published for others to run.

The next things to watch are a technical report or model card with full specifications and scores, the public Physics-IQ Verified listing that would confirm the ranking, and the first hands-on reports from developers who download the weights and code. Results on independent physics tests, such as the IntPhys 2-style possible-versus-impossible checks where current models score near chance, would show whether PAN 2’s physics claims hold up outside its maker’s own evaluations.