PAN 2 vs Hunyuan3D WorldClaw: One Model With Two Halves, or Fifty Models in a Trench Coat

Put these two side by side and the first thing you notice is that they are not the same kind of object. PAN 2 is a model: 29 billion parameters of multi-modal diffusion transformer from the Institute of Foundation Models, Apache-2.0 on Hugging Face since 7 October 2026, code on 5 November, Day-0 vLLM-Omni support on 10 November, and it does one thing — given a starting frame and a described action it renders up to twenty seconds of 704×1248 video at 24 frames per second predicting the consequence. Hunyuan3D WorldClaw is a framework: Tencent Hunyuan's agentic pipeline for turning one open-ended text prompt into an explorable 3D scene made of explicit textured per-object meshes, described in arXiv 2608.05248 on 5 August 2026 and announced on 11 August, with a repository that holds a README and assets and no code. One is a downloadable checkpoint you run. The other is a published method you would have to rebuild.

The second thing you notice is more useful. WorldClaw's own paper names the models it runs on, and that list is a shopping list of other people's components: Claude Opus 4.8 for planning, GPT-Image-2 for imagery, SAM3 and SAM3D for segmentation, Hunyuan3D for reconstruction, Blender 5.1.1 driven over MCP for assembly, all on four NVIDIA H20s. PAN 2 names one component and leaves it swappable: any instruction-following language model writes the structured caption, and IFM ranked four of them. That is the real comparison — not scenes versus video, but what happens to a system when intelligence is spread across a chain of vendors instead of concentrated in one checkpoint.

Two architectures, drawn honestly

PAN 2's design is Generative Latent Prediction from Critique of World Models (arXiv 2507.05169). A conceptual reasoner writes the next world state as a structured, time-grounded description — events in order, actions bound to actors, camera motion — a frozen Qwen3.5-9B text encoder maps it to conditioning features, a 29B diffusion transformer fills in perceptual detail, and a Wan 2.2 VAE decodes the frames. Two halves, one repository, one licence.

WorldClaw's design is a four-stage refinement loop that its authors describe as coarse-to-fine: language-model agents read the prompt and produce a scene plan; a global pass generates semantic layout, height fields and materials so the ground is coherent before anything sits on it; regions are rendered, segmented and reconstructed into individual meshes through Hunyuan3D; and a vision agent inspects the render for floating objects, interpenetration, wrong scale and wrong pose, then sends corrections back. The explicit-geometry promise is kept — you get objects, not a fused blob or a Gaussian splat — and the agents do their work once, before you ever load the scene.

Where the two converge is the part worth internalising: in both systems the pixels are the fast, cheap half and a language model decides what should exist. Where they diverge is who owns the language model. IFM treated the reasoner as a variable and published the harness that ranked the candidates. Tencent treated the reasoner as a dependency and named one.

What the word "agentic" buys, and what it costs

The trap in this comparison is that both announcements use the word, and neither means the same thing by it.

• Agents as authors — WorldClaw's planners decompose a prompt into a scene plan and its vision agent issues corrections; PAN 2's reasoner writes a caption describing what happens next.

• When the intelligence runs — WorldClaw's runs once, in the authoring loop, and is gone when the meshes export; PAN 2's runs once per clip, on every call, for the life of the system.

• Output type — WorldClaw returns a scene file with individual meshes, materials and transforms; PAN 2 returns an MP4 with no geometry in it at all.

• Editability — WorldClaw's output opens in a modeller, gets retopologised and placed; PAN 2's only editable surface is the caption, so revision means re-rendering.

• Dependency surface — WorldClaw requires six separately licensed components and a Blender install; PAN 2 requires the weights, a text encoder and a VAE.

• What "good" means — WorldClaw is judged on whether the scene assembles without floating objects; PAN 2 is judged on whether the consequence it renders is the one the action would actually produce.

The cost of WorldClaw's spread-out intelligence is not conceptual, it is operational: six components means six sets of terms, six failure modes, and six places where a version bump can change the output. The cost of PAN 2's concentrated intelligence is the reverse — one licence, one dependency tree, and a reasoner you have to evaluate yourself because the vendor's ranking of it is the vendor's.

Neither cost is a defect. But it explains why the two are rarely the right answer to the same question: one is a content pipeline, the other is a consequence engine, and a team that needs meshes will find nothing in PAN 2 to help it.

Two pipelines, written as a sequence of stages

Sequence is where the difference between a model and a framework shows up most plainly.

• PAN 2, one call — read the observation and the requested action, write the structured world state, encode it, run fifty diffusion steps at guidance 2.5, decode twenty seconds of 704×1248. Nothing in that sequence persists after the clip.

• WorldClaw, four stages — plan the scene with a language model, generate global semantic layout and height fields, render and segment regions and reconstruct each into a mesh, then inspect the render with a vision agent and send corrections back. Nothing in that sequence is a single model.

• Where the state lives — PAN 2's only durable artefact is the caption and the MP4; WorldClaw's is a scene file with individual meshes, materials and transforms that outlives the pipeline.

• Where it can fail — PAN 2 fails inside the render: a dropped step, drift, a physics shortcut. WorldClaw fails at a seam: a planner that misread the prompt, a segmenter that cut an object in half, a reconstructor that produced the wrong scale, an assembler that interpenetrated two meshes.

• What you can inspect — PAN 2 gives you a video and a token budget; WorldClaw gives you geometry you can measure, and therefore errors you can attribute.

The last row is the one worth keeping. A failure you can attribute is cheaper to fix than a failure you can only see, and a mesh pipeline gives you that for free because the intermediate artefacts are inspectable. A video pipeline gives you a final answer and a caption, and when the answer is wrong your only lever is the caption.

It also explains why WorldClaw's vision-agent correction stage exists at all. Interpenetration, floating objects, wrong pose and wrong scale are the failure modes of assembling geometry from separately generated pieces, and none of them is visible until you render and look. A single model does not need that stage because a single model cannot produce a scene whose parts disagree — it can only produce a scene that is wrong as a whole.

The evaluation gap is not symmetrical, and it points the other way from the licences

WorldClaw's published evaluation is qualitative. The paper compares its output side by side against SynCity, WorldGen, MajutsuCity and Marble, and there is no FID, no CLIP score, no user study with a number attached. For a paper describing a scene-generation framework that is a real gap, and any team reimplementing it should treat "looks better than SynCity" as a hypothesis rather than a result.

PAN 2 published more, and labelled its conditions more precisely than most video-model releases do. Three hundred single-action test cases across six capability categories, 201 image-to-video and 99 text-to-video; one hundred egocentric everyday tasks posed both ways, 200 prompts and 456 steps, scored on step completion and full-plan completion across three seeds; Physics-IQ Verified on the image-to-video setting, normalised so the variance between two real takes equals 100; PhyGround on a five-point scale judged by Qwen3.8-27B; RBench as an unweighted mean over five task and four embodiment categories; and a blind pairwise human study on 140 prompts with thirteen reviewers and 1,400 comparisons, Elo fitted by Bradley–Terry and centred at 1500.

Thirteen reviewers is a small panel, and they were judging the authors' model against the authors' baselines, with the closed comparison models generated once through a third-party inference API and several rows flagged as single-seed. None of it is independent. All of it is legible, which is the next best thing — and it is the difference between a release you can argue with and one you can only look at.

The licences then run the opposite way from the evidence. PAN 2 is Apache-2.0, ungated, 86.17 gigabytes across sixty files, no attribution requirement, no user threshold. WorldClaw has no licence to read, because there is nothing to license: the GitHub project contains a README and assets, no code, no weights, no inference script, no API and no price. A framework paper with no implementation is a description of a method, not a tool, and the four-H20 requirement plus its dependency on a stack of separately licensed models means a release would not be a drop-in download even if it happened.

Where a router is actually load-bearing, and it is not where you expect

This is the rare pair where the interesting commercial fact sits on the framework side. WorldClaw's planning half is not a mystery model behind an API — it is named in the paper, and the components it names are ordinary catalogue models. Claude Opus 4.8 is in OrcaRouter's catalogue. So is GPT-Image-2. So are the four reasoners IFM ranked for PAN 2's caption stage: Claude Opus 5.5, GPT-6 Astra, Kimi K3 and Qwen3.8-Flash.

That means a team reimplementing the WorldClaw decomposition — plan, terrain, objects, refine — can assemble the language and image stages today through one key, at each provider's list price passed through with 0% markup, with automatic failover when a planning call times out mid-build. The same key drives PAN 2's reasoner slot when you swap the caption writer between experiments and hold the 29B renderer fixed on your own node. What OrcaRouter does not do is host either PAN 2 or Hunyuan3D WorldClaw: one is a video diffusion model you run on your hardware, the other does not exist as runnable code anywhere. Nothing in this article should be read as suggesting otherwise.

The honest framing is that we route the seams. In a WorldClaw-style pipeline there are three of them — prompt to plan, plan to image, render to critic — and every one is a metered call. In a PAN 2 pipeline there is one, and it is the only part that does not need a GPU. That is a narrower claim than "we run world models", and it is the true one.

What each design teaches you even if you run neither

WorldClaw's most transferable idea is the decomposition, not the scene. Plan, lay out terrain, reconstruct objects individually, then inspect and correct — that four-stage structure is a sound architecture for any pipeline whose output has to be edited later, and the explicit-mesh choice is the right one for anything that leaves the pipeline and lands in a game engine. The correction loop with a vision agent inspecting for floating objects and wrong scale is the part most teams skip and most teams regret skipping.

PAN 2's most transferable idea is the caption contract. The interface is not a prompt box; it is a structured, time-grounded description with a token budget of 1,480 soft and 1,536 hard, a shipped prompt-rewriter skill, a fixed negative prompt and a validator that checks captions against the encoder's limit. Free text gets rewritten before use, and the model card says plainly that the agent case studies relied on hand-refined prompts. Any team building a world model or consuming one should steal that: the contract between the reasoning half and the rendering half is where quality is won or lost, and it deserves a schema rather than a vibe.

Both designs also share a failure the field has not solved, and they fail at it differently. PAN 2 drifts in appearance across clips because each is simulated from at most one starting frame plus text, and its card names perceptual memory as the missing piece; its own limitation section leads with the admission that about two in three simulated plans contain at least one step that is not carried out. WorldClaw's drift is geometric and spatial: interpenetration, wrong scale, objects that float, which is why it needs a vision agent in the loop at all. One needs a memory, the other needs an inspector, and neither has published a measured curve of how bad it gets over time.

The decision, in one pass

If you need a place — a level, a synthetic interior dataset, a virtual production set — WorldClaw is the more relevant reference and you cannot run it. Take the four-stage decomposition and the explicit-mesh output into your own build, budget the four-H20 class of machine plus a Blender dependency, and replace the qualitative evaluation with something you can measure, because the paper does not give you one.

If you need a consequence — what a plan does, before you commit to it — PAN 2 is downloadable this afternoon, permissively licensed, and honest about failing in two of three plans. Read the limitations page first, keep a checker in the loop, and expect to spend GPU hours rather than dollars per token: thirteen seconds of video takes roughly 119 seconds on four B300s, and the sampler runs a hundred forward passes per clip.

If you are building the thing neither lab shipped — a world you can edit as geometry and then watch behave — you are assembling it yourself out of a refiner, a scene builder, a renderer and a language model at every seam. That is more work than either announcement suggests, and it is where the next year of practical work in this category is going to happen.

What would change this page is code. A WorldClaw release with weights and an inference script would turn a method into a tool and make the comparison a live one; a PAN 2 variant that emitted depth alongside RGB would do the same from the other direction. Neither exists today. Until one does, the accurate summary is a downloadable consequence engine and an unbuildable scene framework, and the only thing they genuinely share is the ambition to predict a world rather than render a picture of one.