Apollo World vs Hunyuan3D WorldClaw: A Place You Can Walk, or a Drive You Can Score

The most useful thing about putting Apollo World against Hunyuan3D WorldClaw is that both use the word "world" and neither means the same thing by it. Hunyuan3D WorldClaw is Tencent Hunyuan's agentic framework for turning one open-ended text prompt into a large, explorable 3D scene made of explicit, textured, per-object meshes that you can open in Blender and export to a game engine. Apollo World is Baidu's agent-driven simulator, disclosed on 3 November 2026, which generates real-time camera and lidar observations of a virtual city so that a robotaxi stack can be trained and graded on complete trips. WorldClaw builds a place. Apollo World runs a drive through one. If you came here to find out which generates better 3D, the question is malformed in a way that is worth unpacking rather than answering.

Two products wearing one noun

The overlap is real but narrow. Both produce something a person could describe as a world. After that, every property diverges.

• Deliverable — WorldClaw: a scene file with individual meshes, materials and transforms. Apollo World: a stream of sensor frames plus agent state, consumed by a driving policy rather than a person.

• Unit of work — WorldClaw: one prompt to one scene. Apollo World: one route, one policy, one graded trip.

• Time — WorldClaw's output is static once generated; Apollo World exists to be run forward, with each decision changing what happens next.

• Who it is for — WorldClaw: a level designer, a game studio, a synthetic-data team that needs geometry. Apollo World: Baidu's own autonomous-driving organisation.

• What "good" means — WorldClaw: does the scene look coherent and assemble without floating objects. Apollo World: does the car under test get safer.

Put that way, the two are not rivals so much as adjacent stages of a pipeline that nobody has yet built end to end. One makes the set. The other directs the traffic on it.

What WorldClaw actually does, stage by stage

WorldClaw is best understood as a conductor rather than an instrument. The paper, arXiv 2608.05248, was submitted on 5 August 2026, with the official announcement following on 11 August. The framework takes a free-form prompt and runs four coarse-to-fine stages.

• Intent and planning — language-model agents read the prompt and decide what the scene should contain, producing a plan rather than an image.

• Terrain — a global pass generates semantic layout, height fields and materials, so the ground is coherent before anything is placed on it.

• Objects — regions are rendered, segmented, and each region reconstructed into an individual mesh through Hunyuan3D. This is where the explicit-geometry promise is kept: the output is a set of objects, not a fused blob or a Gaussian splat.

• Refinement — a vision agent inspects the render for floating objects, interpenetration, wrong scale and wrong pose, and sends corrections back.

The stack underneath is a roster of other people's models and one very large application: Claude Opus 4.8 for planning, GPT-Image-2 for imagery, SAM3 and SAM3D for segmentation, Hunyuan3D for reconstruction, and Blender 5.1.1 driven over MCP for assembly. Tencent reports running it on four NVIDIA H20 GPUs. The published evaluation is qualitative, comparing outputs against SynCity, WorldGen, MajutsuCity and Marble with side-by-side scenes and no FID, CLIP or user-study numbers.

The repository is the part that matters for anyone planning to use it. The official GitHub project contains a README and assets. There is 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 gap between those two things is where most "can I try it" questions die.

What Apollo World actually claims

Apollo World was disclosed at Baidu AI Day·Apollo in Shenzhen on 3 November 2026 by Wang Liang, chief architect of Baidu's Intelligent Driving Group, and Yang Ze, who leads the world model work. The framing was pointed: a world model should not be a painter of impressive video but an examiner of driving safety. Baidu backed that with a rubric it calls "2+3" — two gates that decide whether something counts as a world model at all, and three properties that grade how good it is.

The gates are causality and interactivity. Causality means the next state is predicted from the current one rather than assembled from frames and back-filled; interactivity means an action injected into the scene changes what follows. The graded properties are physical accuracy, persistence and real-time performance. Persistence is the one that separates a demo from a test rig: a simulator whose objects slowly drift or deform over a long run will eventually grade a policy against a world that no longer matches the one it started in. Baidu puts the decision cycle for a road vehicle at one thirtieth of a second, about 33 milliseconds.

Three capabilities carry those promises. Every traffic participant is modelled as an agent with its own decisions rather than a scripted prop, so a braking robotaxi can provoke a different response from the car behind it. Camera and lidar data are generated in real time, which lets the driving stack perceive the virtual city through the same interface it uses on a real road. And trips close end to end — order accepted, passenger collected, passenger delivered — across arbitrary origins and destinations, rather than as isolated scenario clips.

Baidu says the model is already supporting Apollo Go's global testing and validation, and cites roughly 100,000 kilometres of simulated driving per day, framed as about three years of continuous testing by a safety-driver-equipped car. Apollo Go's own operating numbers were quoted alongside: 28 cities, more than 23 million rides, 350 million autonomous kilometres, of which more than 240 million were fully driverless. No parameter count, architecture, benchmark table, paper, weight file or price was published. This is a capability briefing about internal infrastructure, not a product launch.

"Agentic" means two different things here

This is the trap in the comparison, and it is worth slowing down for, because both announcements lean on the same fashionable word.

In WorldClaw, the agents are the builders. Language-model planners decompose a prompt into a scene plan, a vision agent inspects the result and issues corrections. The intelligence is in the authoring loop, and it operates once, before you ever load the scene. Once the meshes are exported, the agents are gone; the world is inert geometry.

In Apollo World, the agents are the inhabitants. The planners are not assembling anything — they are driving the other cars, walking the pavements, deciding whether to yield. The intelligence persists for the duration of the simulation, and its whole purpose is to make the world respond to the thing being tested. A static scene with reactive inhabitants would be a contradiction in WorldClaw's terms and the entire point in Apollo's.

That distinction also explains why the two are hard to combine naively. Dropping WorldClaw meshes into a driving simulator gives you a good-looking set with nothing moving in it. Adding reactive traffic to a WorldClaw scene means writing the behaviour layer yourself. The frameworks solve different halves of the same problem, and neither ships the other half.

Neither is available, and not for the same reason

WorldClaw is documented but not distributable. The paper describes a method in enough detail that a well-resourced team could reimplement it, and several third parties have built partial versions, but Tencent has not released the code or the weights. The four-H20 requirement and the dependence on a stack of separately licensed models also mean a release would not be a drop-in download even if it happened.

Apollo World is deployed but not offered. Baidu describes it as running for Apollo Go, and the disclosure contains no commercial terms, no interface and no indication that external access is planned. The comparison between the two on availability is therefore not "open versus closed" but "unreleased versus internal" — a distinction that matters if you are waiting for either.

For teams that need a world this quarter, the practical route is neither. A game engine with a procedural scene tool covers the static half; a commercial driving simulator covers the dynamic half; and the open-weight video world models cover the demo half, badly, if what you need is sensors. OrcaRouter does not carry Apollo World or Hunyuan3D WorldClaw — neither is a callable model — but the language and vision stages that surround a scene pipeline are ordinary API work, and that is the layer we actually cover: one key, more than 200 models, 0% markup on provider list price, and failover when a planning call times out mid-build.

Which one is the right reference for your problem

If you are building a game level, a synthetic dataset of interiors, or a virtual production set, WorldClaw is the more relevant design even though you cannot run it: the four-stage decomposition — plan, terrain, objects, refinement — is a sound architecture, and the explicit-mesh output is the right choice for anything that has to be edited later. The absence of quantitative evaluation is the caveat to carry into any reimplementation; "looks better than SynCity" is not a specification.

If you are validating a driving stack, Apollo World's 2+3 rubric is the more useful export, and it is usable today as a checklist even though the model is not. Ask any simulator vendor about causality, interactivity, physical accuracy, persistence over a long run and latency under scene complexity. A vendor who cannot answer the persistence question with a drift measurement is selling you a video generator with a physics label.

The two will probably converge. A simulator that could ingest editable, well-structured scenes from a framework like WorldClaw would solve the content bottleneck that makes driving simulation expensive; a WorldClaw-class generator that could animate its inhabitants would stop producing dioramas. Neither company has announced that work. For now the honest description is a set builder with no code and a traffic examiner with no door, and the only thing they genuinely share is a word.