Apollo World vs Hi3D v1.0: One Grades a Drive, the Other Sharpens a Mesh

Apollo World and Hi3D v1.0 arrived eleven days apart, both from Chinese teams, both routinely described as 3D, and that is roughly where the resemblance stops. Apollo World is Baidu's agent-driven world model for robotaxi validation, disclosed on 3 November 2026 at Baidu AI Day·Apollo in Shenzhen: it produces real-time camera and lidar observations of a virtual city in which other vehicles and pedestrians are agents that react to the car under test. Hi3D v1.0 is an open-weights geometry refiner announced on 23 October 2026 that takes a single image plus a coarse mesh and returns a denser one, with no colour, no texture and no notion of time. One outputs a graded sequence of sensor frames; the other outputs one static shape.

That difference decides almost everything a buyer cares about, and it also explains why the two are compared at all. Search for either and you get a page of "3D generation" coverage; the listing pages that rank models by voxel resolution put Hi3D near the top, and the world-model coverage that has followed Baidu since 3 November puts Apollo World in the same paragraph. Neither list is wrong, exactly. Both are answering a question you did not ask.

What each model actually emits

The cleanest way to separate them is to name the output object before naming any capability.

• Input — Apollo World takes a route and a driving policy; Hi3D v1.0 takes one image and a coarse mesh of the same object.

• Output — Apollo World emits sequences of camera and lidar data plus the state of every agent in the scene; Hi3D emits vertex geometry only, re-centred and normalised to unit radius.

• Time — Apollo World is a closed loop that runs forward and feeds the consequence of each decision back to the policy under test; Hi3D has no time axis at all.

• Evaluation — Apollo World is judged on whether a driving stack gets better; Hi3D is judged on geometric fidelity against a reference mesh.

• Access — neither is a self-serve API. Apollo World has no published endpoint, price, weights or parameter count. Hi3D is downloadable and runnable, if you own the right GPU.

Read that list and the matchup collapses into a supply-chain question rather than a head-to-head. Hi3D is the kind of tool that produces a prop. Apollo World is the kind that decides how the prop behaves once a car is moving toward it.

Apollo World is an examiner, not a renderer

Baidu's framing on stage was unusually blunt about what it thinks a world model is for. Wang Liang, chief architect of Baidu's Intelligent Driving Group, said the model should not be a painter producing impressive video but an examiner verifying whether an autonomous driving stack is safe. The company then published a rubric it calls "2+3", and the rubric is the most useful thing to come out of the launch because it tells you what Baidu believes it has built.

The two gates are causality, meaning the future is generated from the present state rather than assembled frame by frame and back-filled, and interactivity, meaning an injected action changes what happens next. A model that fails either is, on Baidu's definition, not a world model. The three graded properties are physical accuracy, persistence — objects and details must not drift over long runs — and real-time performance, where the complexity of the scene must not inflate inference latency. Baidu put the decision cycle for a road vehicle at one thirtieth of a second, roughly 33 milliseconds, which is the number that makes the whole design legible: you cannot serve that budget with a model that samples a video for thirty seconds.

Three capabilities follow from the rubric. Traffic participants are modelled as agents with their own decisions rather than scripted props, so a robotaxi braking or yielding can trigger a different response from the cars around it. The system generates camera and lidar observations in real time, so the driving stack perceives the virtual world through the same interface it uses on a real road. And trips close end to end — arbitrary origin to arbitrary destination, order accepted, passenger carried, passenger dropped — rather than isolated scenario clips spliced together.

Baidu says Apollo World is already in use for Apollo Go's global testing and validation. The operating figures it quoted alongside are Apollo Go's, not the model's: 28 cities, more than 23 million rides, 350 million kilometres of autonomous driving of which more than 240 million were fully driverless. The headline simulation claim is that Apollo World can cover around 100,000 kilometres of virtual road testing a day, which Baidu likens to roughly three years of continuous testing by a safety-driver-equipped vehicle. Treat that as a vendor figure with an unusual amount of arithmetic behind it and no external check at all: the conversion assumes a test car running continuously, and nobody outside Baidu has seen the daily distance it is being compared against.

What Baidu did not publish is equally informative. There is no parameter count, no architecture diagram detailed enough to reimplement, no paper, no benchmark table with named baselines, no weights and no price. Apollo World is infrastructure for Baidu's own robotaxi fleet, described publicly for the first time. It is not a product you can buy, and the company has not said when or whether that changes.

Hi3D v1.0 is a refiner with a narrow, measurable job

Hi3D v1.0, formerly Sparc3D, is the opposite kind of artifact: small enough to download, specific enough to benchmark, and honest about its limits in the technical report. It is not a text-to-3D generator. You give it an image and a rough mesh, and it returns a refined mesh at higher resolution, running three progressive passes at nominal grids of 64³, 128³ and 192³, which extract sparse grids at 512³, 1024³ and 1536³. Training scaled from roughly half a million assets in the Sparc3D paper to about two million curated 3D assets, and the 1536³ level is new.

The output is geometry and only geometry. No texture, no material, no vertex colour. Each result is re-centred and normalised to unit radius, so absolute scale is discarded — a detail that matters if you were hoping to drop the mesh straight into a scene at real-world size. Batch size is one.

The evaluation uses ULIP-2 and Uni3D for semantic consistency plus three metrics the report introduces: multi-view fidelity, detail richness and slicing success rate. Hi3D claims the highest multi-view fidelity and detail richness among the open 3D models it compared against, and a 90% slicing success rate, supported by a 562-judgment user study with 13 participants. Those are the authors' own instruments on the authors' own comparisons, and the repository ships inference code only — no training code, no evaluation harness, no benchmark splits, no comparison meshes. You can check that the model runs; you cannot easily check that it wins.

Practical requirements are steep and worth stating plainly. Linux, Python 3.10, an NVIDIA GPU with compute capability 8.0 or above, CUDA between 12.0 and 12.5, flash-attn with no fallback path, and no CPU inference. The authors validated on a single 96GB H20. The licence is royalty-free for commercial use with a "Powered by Hi3D 1.0" attribution, and a separate agreement kicks in above 200,000 monthly active users or $1 million in monthly revenue — a threshold that tells you the target customer is a studio or a platform, not an individual.

The only place the two genuinely meet

There is a real composition here, and it is worth being precise about it because it is easy to overstate. A driving simulator needs content: streets, kerbs, signage, other vehicles, street furniture. Hi3D is a tool for turning a rough blockout into a detailed mesh of one of those objects. Apollo World is a tool for deciding how those objects and the agents controlling them behave over a trip. In a hypothetical pipeline, Hi3D sits upstream of the simulator, producing geometry that a scene builder then places.

Baidu has not said that Apollo World uses Hi3D, or anything like it, and this article is not going to imply otherwise. What can be said is that the two operate on different axes: Hi3D improves the resolution of a static object, Apollo World improves the realism of a dynamic interaction. Refining a mesh until it has crisp panel gaps tells you nothing about whether a pedestrian will step off the kerb when the car slows. Simulating that pedestrian tells you nothing about whether the mesh you are looking at is the right shape.

Where a routing layer does earn its place is around the pipeline rather than inside it. Both tools are surrounded by language and vision stages — captioning renders, planning scene descriptions, judging whether a generated frame matches the brief — and those stages are ordinary model calls. OrcaRouter carries more than 200 models behind one API at 0% markup, with the provider's list price passed through and automatic failover between providers, so a scene-planning agent and a render critic can share a key and a fallback path without a second contract. Neither Apollo World nor Hi3D v1.0 is in that catalogue, and nothing here suggests otherwise; what we route is the glue, not the simulator.

Cost, access and the honest comparison

There is no price comparison to make, because one side has no price. Apollo World has no published commercial terms of any kind. Hi3D v1.0 is free to download and free to use commercially below the revenue and user thresholds, and the real cost is the hardware: a 96GB accelerator, a CUDA stack pinned to a narrow range, and a dependency on flash-attn that the report says has no fallback. If you do not already own that class of machine, the licence is the cheap part.

That asymmetry is the actual answer to "which should I use". The question assumes a choice between two ways of getting a 3D world. There is no choice. One is a downloadable refiner for people who already have meshes and GPUs. The other is Baidu's internal validation harness, described in public for the first time on 3 November and not offered to anyone outside the company.

How to decide, in one pass

If your problem is that a blockout mesh is too coarse, Hi3D v1.0 is a candidate and Apollo World is irrelevant — there is nothing to call, nothing to license and no output you can consume. If your problem is that your driving stack has not seen enough long-tail interactions, Apollo World is the more interesting design and equally out of reach; the practical response is to read its 2+3 rubric as a specification and hold your own simulator vendor to it, because those five properties are the right questions to ask even when the model that prompted them is not for sale.

If you are building the content layer of a simulator, the two are complementary in the way that a modelling tool and a physics engine are complementary. You would not benchmark them against each other. You would buy the one that exists and design around the one that does not.

What would change this answer is a Baidu announcement that Apollo World is available outside Apollo Go, with a price and an interface. Nothing in the 3 November material points that way; the disclosure was a capability briefing to an industry audience, with the deployment framed entirely around Baidu's own fleet. Until then, the accurate description is not "Apollo World versus Hi3D v1.0". It is a closed simulator and an open refiner that happen to share a fortnight and a keyword.