Hi3D V1.0 vs Odyssey-3: 90% and 66.1 Are Not the Same Kind of Number

The two most quotable figures in open and preview 3D AI right now are 90% and 66.1. The first belongs to Hi3D V1.0, the open-sourced geometry refiner that appeared on 23 October 2026. The second belongs to Odyssey-3, the interactive world model whose public research preview opened on 8 October 2026. Put side by side they look like a scoreboard. They are not comparable, and neither one means what its headline implies — which is the actual reason to read a comparison between these two systems.

Hi3D V1.0 and Odyssey-3 are both described as 3D models and share almost nothing else. One takes an image and a coarse mesh and returns a denser mesh file you can slice, print or import. The other takes a prompt and your camera input and returns a live video of an environment that keeps responding to you. One ships weights under a custom licence; the other ships a browser demo and a contact form. Treating them as alternatives is a mistake worth correcting before it costs anyone a week of evaluation.

What leaves each system

Hi3D outputs a mesh. Specifically, a trimesh.Trimesh object that exports to GLB or OBJ, with geometry only — no textures, no materials, no vertex colours. It arrives recentred and normalised to unit radius, so the real-world size of your input is not preserved and has to be restored downstream. Batch size is one. If you are printing, the file goes to a slicer; if you are modelling, it goes into Blender or a CAD tool. Either way you end the session holding an asset.

Odyssey-3 outputs frames. You prompt it, upload an image if you want, and move a camera through the scene; the model predicts what happens next and renders it. There is no mesh, no export path and no project file. The session ends when you close it. Odyssey describes the system as "a learned dynamical system, implemented as an autoregressive diffusion transformer," which is accurate and also a way of saying that the product is the experience rather than the artefact.

That single difference settles most real evaluations. If a downstream step needs vertices, Odyssey-3 cannot supply them at any price. If the deliverable is an explorable scene, Hi3D produces nothing you can walk through.

Reading the 66.1

Odyssey-3 Pro's 66.10 on Physics-IQ Verified, video-to-video, is the strongest number the company has. It is also a best-of-eight sample. On single generations the same model scores 63.37, and the base Odyssey-3 scores 61.56 against 61.11 for FLUX 3 — a gap of well under a point once you stop cherry-picking. On benchmark-supplied prompts rather than custom ones, Odyssey-3 drops to 51.76, roughly ten points below its headline.

The image-to-video split is a useful sanity check. There, FLUX 3 leads at 54.70 against Odyssey-3 Pro's 54.69. A model that is genuinely state-of-the-art across the board does not lose by a hundredth of a point in an adjacent task; what that near-tie suggests is that the video-to-video lead is real but narrow, and sensitive to how the evaluation is framed.

WorldMark is murkier. Odyssey reports first place in three of four categories, which is true of its own run. Odyssey-3 does not appear in WorldMark's published results, and in the company's own first-person real-environment category it scores 80.6, behind Lyra 2.0 at 84.4 and AlayaWorld at 83.0. A self-run evaluation that places you first in most categories and third in one is not dishonest; it is a vendor benchmark, and it should be read the way all of them are read.

What Odyssey-3 has that Hi3D does not is a third-party leaderboard entry at all. That is worth something. It means at least one set of numbers exists that the company did not compute alone, even if the sampling protocol flatters it.

Reading the 90%

Hi3D's headline is a slicing success rate of about 90%, cited in Chinese-language coverage alongside a comparison to 7–17% for unnamed mainstream models at 1536³. The English release states the figure without the comparison. Neither version names the slicer, the test meshes, the failure criterion or the baseline models.

The repository cannot close that gap. It ships inference code, eight weight files and documentation; it does not ship an evaluation harness, benchmark splits, comparison meshes or a slicer configuration. So the 90% is a claim about a test that cannot be rerun from the release. It might be exactly right. It is not evidence you can act on, and the distinction between those two sentences is the whole of research hygiene.

Slicing is also not printing. A mesh that slices without error can still fail on adhesion, support generation, dimensional tolerance or wall thickness. Hi3D's number, even taken at face value, describes the step before the one that actually fails in a workshop. The companion claim — multi-view fidelity 0.75, detail richness 0.841, an 85.9% human-preference score — is likewise vendor-reported with no independent replication.

So the two headline numbers are uncheckable in different ways. Odyssey-3's 66.1 is measured by someone else but sampled generously. Hi3D's 90% is sampled however the vendor liked and measured by nobody else. Neither is a lie. Both are weaker than the sentence that carries them.

The access asymmetry is larger than the quality gap

Hi3D released weights. Eight files on Hugging Face, inference code on GitHub, a free demo Space, and a custom licence that permits commercial use below 200,000 monthly active users or $1,000,000 in monthly revenue. Above those thresholds you must notify Hi3D and negotiate a separate licence within 60 days, and any integration must display "Powered by Hi3D 1.0." It is not an OSI-approved licence, and calling it open source without that caveat is imprecise.

Odyssey-3 released no weights, no code and no licence. Access is a free browser preview at experience.odyssey.systems, plus API access by request. The developer page currently advertises Odyssey-2 Max rather than Odyssey-3, and one preview button refers to an "Odyssey-3 Flash" tier the company does not otherwise document. No price has been published — no rate card, no per-second charge, no credits. The only financial figure in circulation is an assumed $1 per MI355X GPU-hour used inside a cost calculation.

For a team that needs to embed a model in a product, that asymmetry decides the question before quality enters it. Hi3D can be pinned, quantised, run air-gapped and inspected. Odyssey-3 can be called, if you are accepted, at a price you will learn later.

What each costs to actually run

Hi3D's floor is steep and specific. Linux only, Python 3.10, an NVIDIA GPU at compute capability 8.0 or above, CUDA 12.0–12.5, and flash-attn with no fallback. There is no CPU path. Hi3D validated 1536³ on a single 96 GB H20, and the repository does not claim a minimum VRAM below that. Two extra large vision encoders — DINOv2-large and CLIP-ViT-L/14 — are fetched separately. You pay in hardware and setup time, and the marginal cost of a run is electricity.

Odyssey-3's floor is a Chromium browser. That is the appeal and also the limit: you cannot buy capacity, only request access. The company has published generation-cost charts rather than serving prices, so the honest statement is that the cost of embedding Odyssey-3 in a product is currently unknown. A preview is not a price.

The comparison that matters for budgeting is therefore not 90% versus 66.1. It is "a capital expense you can size today" versus "a commercial term that does not exist yet."

Three questions that decide it

• Do you need a file or an experience? If anything downstream needs vertices — a slicer, a game engine, a CAD import — the answer is Hi3D and the rest of this section is moot. If the deliverable is an explorable environment, Odyssey-3 is the only one of the two that produces one.

• Do you need to own the model? Regulated data, air-gapped deployment, fine-tuning and reproducible versioning all point to Hi3D. A hosted preview with no published terms cannot satisfy any of them. If you are prototyping an experience and the data is not sensitive, the browser preview costs nothing to try.

• What can you verify before committing? With Hi3D you can run the model on your own meshes this week and form a private opinion. With Odyssey-3 you cannot run it on your own terms at all until access is granted, and even then the benchmark you would want — an independent replication of the 66.1 — does not exist.

Where OrcaRouter fits

Neither model is served through our API, and this comparison should not be read as an offer to call either one here. What we do route is the multimodal work that surrounds them: reviewing renders against a reference photo, describing defects in a turntable, drafting the QA report. One key across 200-plus models, with automatic failover when a provider degrades, keeps that loop independent of whichever 3D tool you settle on.

There is a practical reason to care about that separation. Both of these systems are moving targets — Hi3D has shipped an inference-only repo with no eval harness, and Odyssey-3 is a preview with undocumented tiers. A review pipeline that depends on a single vendor is fragile in exactly the way the underlying models are. Routing the judgement layer separately is the cheapest insurance in the stack.

One clarification worth having: Odyssey's Agora-2, the multi-agent playable world preview, is a separate product from Odyssey-3, and comparisons written against Agora-2 do not transfer. The numbers above are Odyssey-3's.

What to watch next

For Hi3D, the release that would matter is not a better model but a better artefact: an evaluation harness with the test meshes and slicer settings attached. Until that exists, every quality claim about V1.0 is a vendor statement, and the useful work is your own twenty-object test.

For Odyssey-3, the missing pieces are a published price, a documented tier list and an independent replication of Physics-IQ. The gap between 66.10 best-of-eight and 51.76 on benchmark prompts is the number to watch; if that narrows in a future release, the model is getting more robust rather than better at the test.

Until either happens, the honest verdict is that these two systems do not compete. One sells you a mesh and a GPU bill. The other sells you a view of a world it will not let you keep.