NewsAdGraam raises ₹3 crore in seed funding
AI Video & Image News

PAN 2: 29B Open-Weights World Model, Apache 2.0

AdGraam News Desk5 min read
A strawberry splashing into a glass bowl of liquid in a dark studio, evoking a physics simulation world model

The short answer

The Institute of Foundation Models (IFM) at MBZUAI announced PAN 2 on X on November 10, 2026, calling it an open foundation world model for physical intelligence. According to OrcaRouter's write-up, it is a 29-billion-parameter video diffusion transformer with Apache-2.0 weights on Hugging Face, text-to-video and image-to-video support in vLLM-Omni, and a 704x1248, 24 fps backbone. No hosted endpoint or pricing has been announced.

On this page (11 sections)

What is PAN 2?

PAN 2 is an open foundation world model from the Institute of Foundation Models (IFM) at MBZUAI. According to OrcaRouter, it is a 29-billion-parameter video diffusion transformer. IFM says it predicts what happens after each action, grounded in time, physics and space, instead of only turning a prompt into a clip.

IFM's own post on X reads: "A language model predicts the next word. PAN 2 predicts the next world." The lab frames the goal as AI that simulates "what if?" before it acts.

A post from @kimmonismus, who tracks AI releases, describes a demo: drop a strawberry into a bowl, then swap it for a rubber duck or change the liquid, and PAN 2 simulates each version and shows how the splash changes. Creator @AngryTomtweets says what convinced him was a side-by-side with real footage and the open weights.

PAN 2 release date and availability

IFM posted the announcement from @IFM_AI at 17:00 UTC on November 10, 2026. According to OrcaRouter, the weights are already public on Hugging Face under Apache-2.0, and a serving path in vLLM-Omni merged the same day.

The order of events was backwards for a launch. According to a write-up on OrcaRouter's blog, the Hugging Face repository dates from October 7 and the GitHub repository from November 5. It also reports that vLLM-Omni's Day-0 pull request for text-to-video and image-to-video merged at 16:03 UTC on November 10, and the recipe page followed at 20:55 UTC.

That write-up says the announcement carried a video but linked to no blog, paper, hosted endpoint or pricing page. IFM's post promises "more details" through a link, but nothing further is in hand yet.

PAN 2 key features and specs

According to OrcaRouter, PAN 2 is a diffusion transformer with 29 billion parameters and supports text-to-video and image-to-video. A Reddit listing of the Hugging Face repo gives the diffusion backbone as 704x1248 at 24 fps. OrcaRouter lists the checkpoint at 86.17 GB across 60 files, ungated, with inference and training code on GitHub.

Server hardware with stacked translucent data slabs representing a large open model checkpointServer hardware with stacked translucent data slabs representing a large open model checkpoint

SpecDetails
DeveloperInstitute of Foundation Models (IFM), MBZUAI
Model typeVideo diffusion transformer, world model
Parameters29B (per OrcaRouter)
Resolution704x1248 (per Reddit listing)
Frame rate24 fps (per Reddit listing)
ModesText-to-video, image-to-video
Max durationNot announced
AudioNot announced
Weights86.17 GB, 60 files, ungated (per OrcaRouter)
LicenseApache-2.0 (per OrcaRouter)
Hosted API / pricingNot announced

The 704x1248 frame is close to a 9:16 vertical ratio, which matches Reels, Stories and TikTok. IFM has not said how long a clip can run.

The GitHub repository, per the same write-up, includes inference scripts, environment pins, training configs, a prompt-rewriter skill and deployment documentation. A commit on November 9 reportedly added safety-checker support as a Diffusers modular pipeline component.

PAN 2 pricing and free access

PAN 2 has no per-second or per-image price because no hosted endpoint has been announced. If the Apache-2.0 license holds as OrcaRouter reports, the weights are free and the cost is the GPU time you run yourself.

OrcaRouter lists the checkpoint at 86.17 GB, which is far beyond a laptop download. Whether a hosted version is coming is not announced.

PAN 2 vs other AI video models

IFM has not published ad-oriented benchmarks or comparisons. No head-to-head numbers against commercial video models are available, so treat any quality ranking as unproven.

What differs is the positioning. Commercial video models sell clip quality, duration and audio. PAN 2 is pitched as a world model that responds to actions, with open weights and training code. For comparison with a recent unconfirmed model story, see BouncyBohr: Leaked Google Video Model, 4K Claims.

@AngryTomtweets points to a side-by-side with real footage. Check that comparison yourself before trusting it. Until a paper or model card appears, claims about physics fidelity rest on IFM's own demo video.

How to use PAN 2 for AI video and static ads

PAN 2 is an open model with weights and code, and no hosted ad product has been announced. Its near-term fit is product physics shots and b-roll tests, if you can run it yourself. Fit for hooks, UGC or localization is unproven, and no audio is announced.

A creative team desk with a vertical phone and storyboard cards for planning product ad shotsA creative team desk with a vertical phone and storyboard cards for planning product ad shots

Where it might help, if outputs hold up:

  • Product interaction b-roll: splashes, pours, drops and object swaps, the exact demo IFM and @kimmonismus show.
  • Vertical-first tests: the listed 704x1248 backbone is close to 9:16, though final delivery resolution for paid social is not announced, so you may need an upscaling step.
  • Variant exploration: the "change the liquid, change the object" framing maps to testing product variants before a shoot.

Practical limits: no audio, no confirmed clip length, no hosted API. A creative team would want PAN 2 outputs chained with upscaling and editing steps, which is the kind of repeatable setup AdGraam Flow Builder handles for supported models. You can compare other video models side by side in the AdGraam Playground.

What to watch next

Watch for IFM's promised details page, a paper or model card, benchmark numbers against real footage, and any hosted endpoint with pricing. Until those appear, ad teams have little to evaluate beyond the open checkpoint itself.

Also watch for clip duration and audio support, which are the two specs that decide ad usability. Outputs from people running it through vLLM-Omni would add evidence. For now, the only footage is IFM's own video and the demo described by third parties.

From the timeline

What people are posting

Drop a strawberry into a bowl. Now make it a rubber duck. Or change the liquid. PAN 2, the world model the Institute of Foundation Models (IFM) just launched, simulates each version and shows how the splash changes. IFM built PAN 2 to be grounded in time, physics, and space. https://t.co/vGjfpI6MiJ
@kimmonismus8 likesView on X
Skeptical of world-model releases by default. What moved me on PAN 2: a side-by-side with real footage, and open weights so I can test the rest myself.
@AngryTomtweets3 likesView on X

Key takeaways

  • IFM announced PAN 2 on X at 17:00 UTC on November 10, 2026, as an open foundation world model for physical intelligence.
  • According to OrcaRouter's write-up, PAN 2 is a 29-billion-parameter video diffusion transformer with 86.17 GB of ungated Apache-2.0 weights across 60 files on Hugging Face.
  • A Reddit listing of the Hugging Face repo gives the diffusion backbone as 704x1248 at 24 fps.
  • OrcaRouter reports that vLLM-Omni merged Day-0 text-to-video and image-to-video support for PAN 2 on November 10, 2026.
  • No hosted API, pricing page, paper or blog post has been published yet.
  • Clip duration and audio support are not announced.
Run today's top models

Compare 200+ AI video and image models side by side

Run Kling, Veo, Sora, Seedance, Flux, Ideogram and Recraft in one Playground, or turn any product URL into ready-to-launch ads.

Questions, answered

Frequently asked

When was PAN 2 released?

IFM posted the announcement on X at 17:00 UTC on November 10, 2026. According to OrcaRouter, the Hugging Face repository was created October 7 and the GitHub repository on November 5, and vLLM-Omni support merged the same day as the announcement.

How much does PAN 2 cost?

According to OrcaRouter, the weights are Apache-2.0, so you would pay only for your own compute. No hosted endpoint or per-second pricing has been announced.

Is PAN 2 free to try?

OrcaRouter reports the checkpoint is downloadable from the IFM/PAN2 repository on Hugging Face without a gate. It is listed at 86.17 GB, so running it needs serious GPU capacity. No free web demo is mentioned.

How do I get access to PAN 2?

Download the weights from Hugging Face or clone the github.com/ifm-ai/PAN2 repository, which OrcaRouter says holds inference scripts and training configs. OrcaRouter also reports a vLLM recipe page for serving text-to-video and image-to-video.

What is PAN 2 for?

IFM describes it as a world model that predicts what happens after each action, grounded in time, physics and space. Its demos so far show physical interactions such as a splash changing with the object or liquid.

Can you use PAN 2 for commercial ads?

OrcaRouter reports an Apache-2.0 license, which generally permits commercial use. Check the license file and model card yourself, and note that a safety-checker component was reportedly added on November 9.

Where this comes from

Sources

  1. PAN 2 Shipped Quietly, Then Got a Serving Path Firstorcarouter.ai
  2. IFM/PAN2 · Hugging Face : r/LocalLLaMAreddit.com

This story is still unconfirmed: it is based on what people are posting, and the company behind it has not announced it yet. We will update this page as it develops.

AI Video & Image Newspan 2pan 2 world modelifm mbzuai pan 2open weights video modelapache 2.0 video modelworld model video generationvllm-omni pan 2pan 2 specs
Keep reading

More AI model news & creative guides

Your move

See your product as a winning ad

Turn a product URL into static ads and UGC-style videos, rebuild the competitor ads that already work, and read your Meta data to know what to make next.

AdGraam News DeskAI video and image news

The AdGraam News Desk (part of AdGraam, also searched as Adgram or Adgram AI) tracks AI image and video model launches, leaks and rumors from labs, leakers and creators, and checks each claim against its sources. Unconfirmed stories say so, and each page is updated as its story develops.

LinkedIn