Hi3D v1.0 vs Agora-2: A Thing You Can Print vs a World You Can Walk In

Hi3D v1.0 and Agora-2 get filed next to each other because both are described as 3D models, and that filing is wrong in a way that matters. Hi3D v1.0, open-sourced on 23 October 2026, returns a mesh file you can slice, print, or import into a game engine. Agora-2, Odyssey's playable research preview published on 21 September 2026, returns a session: a shared, real-time world that four human players and sixteen AI agents move around in, and that stops existing when everyone logs off. One is an asset pipeline. The other is an experience. The only thing they share is a word.

That is not a reason to skip the comparison. It is the comparison. If you are choosing between them, you are really asking which stage of a production pipeline you are standing at, and the two answers do not overlap.

What actually leaves the machine

Start with the artifact, because it settles most of the argument. Hi3D v1.0 produces a trimesh.Trimesh — geometry only, recentered and normalized to unit maximum radius, with no textures, materials or vertex colours. It is a file. You can open it in Blender, repair it, scale it, and send it to a slicer or an engine. Its limitations are all about what is missing from the file, not about whether the file exists.

Agora-2 produces frames. A learned simulation model tracks world state and a roughly one-billion-parameter flow-matching renderer draws what each participant sees, at 640×480 per player and a target of about 30 frames per second. There is no mesh, no depth buffer, no coordinate system, and no export path. The technical report describes a research preview, not a product with an asset format. When the session ends, the world goes with it.

So the first question is not "which is better." It is "what do you have to hand to someone else at the end of the day?" If the answer is a file, Agora-2 is not a candidate, however impressive the demo looks.

What goes in

The inputs are conditioned on opposite ends of the problem.

• Hi3D v1.0 — a reference image and a coarse mesh of the same object. Not a single photograph. The repo ships an optional adapter that uses Tencent's Hunyuan3D-2 mini to build that coarse mesh, which is where the actual image-to-geometry step happens. Hi3D's job is to add detail to structure it did not invent.

• Agora-2 — player actions. The model was trained on captures of Diablo II gameplay, pairing observations with actions and state, and it renders the consequences of what participants do. There is no object to upload and no mesh to refine. The input is behaviour.

That asymmetry explains the outputs. Hi3D is conditioned on an object, so it can return an object. Agora-2 is conditioned on actions, so it returns the visible result of actions, which is a scene that changes.

Speed, scale, and what real time costs

• Hi3D v1.0 — offline and slow by design. One pass at 512³, two at 1024³, three at 1536³, with 25 scheduler updates per pass. The validated configuration is a 96 GB NVIDIA H20, batch size one, no CPU path. There is no published latency table, and the vendor has not claimed one.

• Agora-2 — interactive. A four-player session runs on four NVIDIA RTX PRO 4500 GPUs, one renderer per GPU. The simulation model is small — roughly 4.46 million parameters — while the renderer carries the visual load. Twenty participants share one world.

The trade is legible: Hi3D spends minutes of GPU time to buy geometric resolution; Agora-2 spends four GPUs to buy liveness at 640×480. Neither is a worse use of hardware. They are buying different things. One caveat worth stating plainly: Odyssey has not published sustained frame rate or input-to-display latency for multi-agent play, and the published comparisons are internal ablations against its own configurations.

Openness, on one axis

This is where the gap stops being philosophical.

• Hi3D v1.0 — downloadable. Eight weight files, inference code, Docker constraints and documentation on GitHub and Hugging Face, under the Hi3D Open License Agreement 1.0. Commercial use is permitted; passing 200,000 monthly active users or USD 1,000,000 in monthly revenue triggers a notification and a separate written licence within 60 days, and integrations must carry a "Powered by Hi3D 1.0" credit. The Hugging Face card notes the model "isn't deployed by any Inference Provider," so self-hosting is the only route.

• Agora-2 — not downloadable. No weights, no repository, no licence, no public API, no pricing. What exists is a technical report and a browser demo. The only way to use it is Odyssey's hosted preview.

A restrictive licence on a runnable model still beats a permissive-sounding demo you cannot run, if the goal is to build something. Hi3D's terms are a diligence item, not a wall. Agora-2 has no terms to diligence because there is nothing to license.

The stage neither of them covers

Put the two side by side and a hole appears in the middle. Hi3D cannot invent structure: it needs a coarse mesh, and the README warns that severe coarse-mesh errors can survive refinement. Agora-2 cannot produce an object at all. The step that turns one photograph into a plausible coarse mesh — the actual first draft of geometry — belongs to neither release. It belongs to Hunyuan3D-2 mini, TripoSR, or whatever your studio already uses.

That is the honest map of this corner of the market in October 2026: the world-simulation end is being demonstrated behind a hosted preview, the geometry-refinement end has been opened up, and the bootstrap step in between is still a separate tool with its own failure modes.

OrcaRouter sits outside all three. We do not host Hi3D v1.0, Agora-2, or a coarse-mesh generator, and nothing here is a route to a 3D asset. What we do serve is the surrounding text and vision work — expanding a brief into generation prompts, captioning a reference image, or running a vision model over multi-view renders to catch defects before they reach a slicer. That is one API for more than 200 models behind a single key, with automatic failover when a provider degrades. It is the part of the pipeline that is cheaper to rent than to build, which is exactly why it is worth naming separately from the part that has to run on your own GPU.

Which one, for which job

• You need a physical object — a part to print, a prop to manufacture, a mesh to hand to a fabricator. Hi3D v1.0, with the scale, topology and texture caveats attached. The 90% slicing figure in its press release is vendor-reported and not reproducible from the repo, so plan your own test prints.

• You need a place — a prototype environment, a training ground for agents, a demo where the world itself is the product. Agora-2 is the more interesting research artifact, but it is a preview you cannot ship, host, or fine-tune. Treat it as a reference design, not a dependency.

• You are researching multi-agent world models — Agora-2's report is the relevant document, and the absence of weights is itself a finding about where the field's openness currently stops.

• You are finishing assets at volume — Hi3D v1.0, if you have H20-class hardware or are willing to validate the smaller 512³ path yourself. There is no published memory table below the validated card, so that validation is on you.

The comparison that would actually be useful is not Hi3D v1.0 versus Agora-2. It is Hi3D v1.0 versus the coarse mesh you fed it — a before-and-after that shows whether the refinement earned its GPU time. Nobody has published that yet either. Until someone does, the safe reading is that these two releases are adjacent, not equivalent, and that the word "3D" is doing more marketing work than technical work.