PAN 2 Got a Serving Path Before It Got an Announcement

At 16:03 UTC on 10 November 2026, a pull request adding Day-0 support for PAN 2 text-to-video and image-to-video generation finished merging into vLLM-Omni. At 20:55 UTC the same day, a second pull request added the recipe page that tells you how to serve it. PAN 2 is a 29-billion-parameter video diffusion transformer from the Institute of Foundation Models at MBZUAI, and the reason those two merges are the story is the order they arrived in. The Hugging Face repository dates from 7 October 2026 and the GitHub repository from 5 November, so for five weeks the only public trace of the model was repository metadata. The first first-party statement I can find is a single post from the lab's own X account, @IFM_AI, timestamped 17:00 UTC on 10 November — fifty-six minutes after the vLLM-Omni merge and nearly four hours before the recipe. It reads "Introducing PAN 2: an open foundation world model for physical intelligence", carries a video, and links to nothing beyond itself. No blog entry, no paper URL, no hosted endpoint, no pricing page, no press release. So the serving path really did precede the announcement, and the announcement, when it came, was smaller than the merge: a few hundred likes and one pair of paragraphs. What exists is 86.17 GB of Apache-2.0 checkpoint, a framework integration, and a tweet.

What actually landed, and in what order

Four separate things happened, and conflating them is how a quiet release gets mistaken for a launch.

• The weights — IFM/PAN2 on Hugging Face, repository created 7 October 2026 at 23:42 UTC, last modified 9 November 2026 at 21:52 UTC with a commit adding safety-checker support as a Diffusers modular pipeline component. Sixty files, 86.17 GB, ungated, Apache-2.0.

• The code — github.com/ifm-ai/PAN2, created 5 November 2026, with a single commit pushed at 20:13 UTC that day. Not a research dump with a dead README: there are inference scripts, environment pins, training configs, a prompt-rewriter skill and documentation for deployment, training and inference.

• The serving path — vLLM-Omni PR #8602 for Day-0 text-to-video and image-to-video support, merged 10 November at 16:03 UTC, and the vLLM recipe PR #10126, merged the same evening at 20:55 UTC.

• The statement — @IFM_AI on X, 10 November at 17:00 UTC: "Introducing PAN 2: an open foundation world model for physical intelligence." Two paragraphs, a video, no link to a post, a paper or a page.

That sequencing is unusual, and the announcement is the item that makes it stranger rather than more normal. Normally a lab announces, publishes weights, and framework support trickles in over the following weeks while users file issues. Here the framework merge came first, the announcement came fifty-six minutes later, and the recipe came four hours after that. The most substantial public artifact in the model's short life is still a pull request, which tells you something about how this release was run: as an artifact drop for people who read repositories, with a social post appended almost as an afterthought.

What the repository will tell you

More than most release posts. PAN 2 is a world model, not a video generator with a cinematic bent — the distinction is architectural rather than rhetorical. It implements the Generative Latent Prediction pattern described in Critique of World Models (arXiv 2507.05169) and splits the job in two. A conceptual reasoner, which is any instruction-following language model, reads the observation and the requested actions and writes the next world state as a structured, time-grounded caption: events in order, actions bound to actors, camera motion. A frozen text encoder maps that caption into conditioning features. Then a 29B multi-modal diffusion transformer, trained from scratch on hundreds of millions of videos annotated in the same format, predicts the perceptual detail the caption leaves open. A Wan 2.2 VAE decodes the latents, and what you get is a video of up to 20 seconds at 704×1248 and 24 frames per second.

The published configuration is specific enough to cost out. The transformer config lists 44 layers, 32 attention heads at 128 dimensions each, a patch size of 1×2×2, 97 input channels and 48 output channels. The default sampler is 50 steps at a flow shift of 9 with classifier-free guidance of 2.5 on every step. Default length is 241 frames, about ten seconds; the ceiling is 481 frames. The text encoder is Qwen3.5-9B, frozen and unmodified, and the VAE is Wan 2.2's. Both are downloaded from their own repositories — MBZUAI does not redistribute them, and the model card is careful to say so.

There is also a prompt-rewriter skill shipped in the repo, with a system prompt, a fixed negative prompt and a validator that checks captions against the text encoder's token budget: a soft limit of 1,480 tokens and a hard capacity of 1,536. That is the most useful single artifact in the drop. It tells you the model does not really take free text well, and that the team knows it.

What it will not tell you

Almost everything a buyer would ask. There is no price, because there is no API. There is no hosted endpoint from MBZUAI. There is no independent benchmark: the numbers in the model card are the team's own, measured on their own harness, and the card is unusually candid about the conditions — Physics-IQ Verified normalized so that the variance between two real takes equals 100, PhyGround scored on a five-point scale by Qwen3.8-27B, RBench as an unweighted mean over five task and four embodiment categories. Those are defensible choices, and none of them are third-party verification.

The human evaluation is the strongest evidence in the drop and also the most limited. Blind pairwise preference over 140 prompts — 90 text-to-video, 50 image-to-video — judged on action adherence, physics and world consistency by 13 reviewers across 1,400 comparisons, with Elo fitted by Bradley–Terry and centred at 1500. Thirteen reviewers is a small panel and they were judging the team's own model against baselines. It is more than most labs publish. It is not a leaderboard.

And the limitations section is blunter than the genre norm: about two in three simulated plans contain at least one step that is not carried out. That single sentence is the most important number in the release, and it is the one a launch post would have buried.

The throughput figure in the recipe is the real story

Framework support matters because of what the recipe claims, not because a logo was added. The vLLM recipe page states that PAN 2 generates 13 seconds of 704×1248 video in roughly 119 seconds on four B300 GPUs, using classifier-free-guidance parallelism across two ranks and Ulysses sequence parallelism across two. The default configuration is eight H200s in BF16, and the documented serve command adds VAE patch parallelism across four ranks in tile mode — a detail that only matters because decoding a 20-second 704×1248 clip is its own memory problem.

Read those numbers against the sampler settings. Fifty steps with guidance on every step means a hundred forward passes through a 29B transformer per clip. That is why the guide runs guidance in parallel rather than sequentially. It is also why the practical question about PAN 2 is not "how good is it" but "what does a clip cost me", and the answer on the vendor's own recipe is a four-GPU node for two minutes per thirteen seconds of video. Nobody has published a cheaper configuration, and the recipe's eight-H200 default suggests the four-B300 setup is a squeeze rather than a comfortable floor.

If you want to try it, the planner is the part you can rent

PAN 2's two-stage design has an odd practical consequence. The expensive half — the 29B diffusion backbone — is the half you have to self-host, because no one offers it as a service and MBZUAI has not built one. The cheap half, the conceptual reasoner that writes the structured world-state caption, is an ordinary text model and can be anything instruction-following.

That is where a router earns its place. The PAN 2 report evaluates four reasoners and ranks them by full-plan completion: Claude Opus 5.5 first, then GPT-6 Astra at medium effort, then Kimi K3, then Qwen3.8-Flash. Qwen3.8-Flash is on OrcaRouter, so you can drive PAN 2's caption stage through one API key with automatic failover rather than standing up a second local model next to the diffusion stack, and pay the provider's list price with 0% markup passed through. The renderer stays on your GPUs; the planner does not have to.

Be clear about what that is and is not. OrcaRouter does not host PAN 2 and cannot route it — there is no hosted PAN 2 anywhere to route to. What it can do is keep the language half of a world-model pipeline on a metered API while you size the video half against your own hardware. If you would rather run both stages locally, the repo supports that too, and the prompt-rewriter script is included for exactly that case.

What would count as an announcement

Three things, in ascending order of significance. A post, paper or page from IFM or MBZUAI that names PAN 2 and states a release date or a version — the 10 November tweet cleared the first half of that and not the second. A hosted inference endpoint, which would make cost comparison possible against closed video models rather than against GPU rental. And an independent benchmark run by someone who is not the model's author — Physics-IQ, PhyGround or RBench scored by a third party, on prompts the team did not choose.

Until at least the first of those arrives in full, the honest description of PAN 2 is this: an Apache-2.0 research artifact with a working serving path, a documented prompt format, a candid limitations section, one social post and no commercial offer behind it. That is more than a preprint and less than a product. The framework support landing on 10 November did not make it a launch. It made it runnable by people who do not want to patch a repository to get a first frame.

Bottom line

PAN 2 is the rare case where the engineering is ahead of the paperwork. If you are building agent simulation, robotics evaluation or physical-reasoning research and you have a four-GPU node, the weights are there, the recipe is there, and the two-stage design is documented well enough to reproduce. If you need a price per clip, a service-level agreement or a vendor to call, none of that exists yet, and the 10 November merges did not create it. The model shipped quietly, if "shipped" is even the word for a repository, a merge and a two-paragraph post. Whether it ever gets announced in the sense a buyer would recognise is a separate question from whether it works, and only one of those has an answer today.