Genie 4 vs Hi3D v1.0: A World You Visit or a Mesh You Keep

Genie 4 hands you an experience that ends when the session does. Hi3D v1.0 hands you a file. That is the whole comparison, and it is worth stating plainly because the two models get discussed as if they were adjacent: both are generative, both produce three-dimensional-looking output, and both arrived within a week of each other in October 2026. Genie 4, announced by Google DeepMind on October 27, is a real-time interactive world model that generates video with synchronized audio from a text prompt and lets you steer it while it runs. Hi3D v1.0, released on October 23, is an open-weights geometry refiner that turns a coarse mesh plus a reference image into a high-resolution mesh and stops there — no texture, no animation, no scene, no sound.

Read the launch materials side by side and the mismatch becomes obvious in a useful way. Google's post is about what you can do inside a generated world; Hi3D's model card is about how many vertices the refiner can resolve. One is a product experience with no export. The other is an asset tool with no experience. If you are choosing between them, you are not choosing between two world models — you are choosing between two deliverables, and only one of them can be put in a file.

What Hi3D v1.0 actually ships

The repository is unusually explicit about scope, which makes it easy to evaluate. Hi3D v1.0 refines a coarse mesh into a high-resolution one, conditioned on a reference image of the same object. It is built on the Sparc3D representation and refinement architecture: sparse latents are sampled on coordinates taken from the coarse mesh, decoded by a self-pruning sparse variational autoencoder, and the surface is extracted, revoxelized and repeated. The published effective sparse extraction grid goes up to 1536 cubed. The project report describes training on roughly two million curated 3D assets.

What you get is geometry, and only geometry. There is no texture, no material, no rig, no animation and no scene placement. The output is a mesh that has been recentered and normalized to a unit maximum radius, which is a deliberate choice for a refiner and an inconvenient one for anyone who wanted a metric-scale asset. The model card says the training code and the paper's evaluation harness are not included; the inference code is in the GitHub repository, and the weights are on Hugging Face under the Hi3D Open License 1.0.

The compute profile is the part that decides whether this is a weekend project or a procurement request. The documented path is a 96 GB NVIDIA H20 with no CPU fallback and batch size one. That is a data-center accelerator, not a consumer card, and it puts the model in the same hardware class as a small fine-tuning job. There is also no published numeric benchmark — no Chamfer distance, no F-score, no comparison table against Sparc3D or Hunyuan3D. The repo asks you to trust the resolution and the training-asset count, both of which are the project's own figures.

What Genie 4 actually ships

Genie 4 ships a session. You prompt a world, steer it in natural language — move the camera, change the weather, add a character — and it generates frames with native audio in real time, keeping the scene coherent for a reported two minutes. Gemini Omni anchors appearance and behavior, and Gemini 3.8 Flash sits in the loop checking frames and re-planning, which is what makes the model usable as a task-rehearsal environment rather than a clip generator.

The output is pixels and sound, consumed live. There is no mesh, no point cloud, no depth map, no camera track and no export path into Blender, Unity or a slicer. Google does not publish resolution, frame rate, model size, benchmark or price, and access is a limited trusted-tester pool with a waitlist; AI Ultra subscribers through Project Genie and Gemini API early-access partners are promised "soon" without a date. The failure modes Google admits — object permanence drift, imperfect spatial accuracy, illegible in-world text, unsolved audio accuracy — all describe a system whose value is in the moment of interaction, not in the artifact afterward.

Dimension by dimension

• Deliverable — Genie 4 produces a live video-and-audio world; Hi3D v1.0 produces a static mesh file.

• Persistence — Genie 4's world ends when the session ends; Hi3D v1.0's mesh can be stored, versioned and shipped.

• Input — Genie 4 takes text plus steering instructions; Hi3D v1.0 takes a reference image and a coarse mesh, which means it cannot generate from text alone.

• Surface treatment — Genie 4's output is fully shaded frames; Hi3D v1.0's output is untextured geometry, with appearance left to a downstream texturing step.

• Compute — Genie 4 runs on Google's infrastructure behind an invitation; Hi3D v1.0 runs on a 96 GB H20 that you have to own or rent.

• Access — Genie 4 is waitlist-only with no price; Hi3D v1.0's weights are downloadable now under the Hi3D Open License 1.0, with commercial terms that trigger above a user and revenue threshold.

• Evidence — Genie 4 has demos and vendor claims with no published benchmark; Hi3D v1.0 has a project report and resolution figures with no independent 3D benchmark either.

The overlap people assume, and the gap that is actually there

The assumption is that a world model and a 3D generator are two routes to the same asset. They are not, for a structural reason: Genie 4 generates frames autoregressively and does not maintain an explicit 3D representation. Its consistency is an emergent property of the sequence, not a scene graph you can inspect. That is why the world drifts after repeated edits, and it is also why there is nothing to export. A viewer sees a three-dimensional place; the model has not built one.

Hi3D sits on the other side of that line. It assumes someone already has a coarse mesh — from a generator, a scan or a CAD approximation — and makes that mesh sharper. It cannot invent an object from a prompt, cannot place two objects in a room, and cannot animate anything. Feed it a bad silhouette and it will return a sharper bad silhouette, which is a real risk the model card does not discuss.

So the formats do not meet in the middle. There is no published pipeline that takes a Genie 4 session and turns it into a mesh, and there is no published pipeline that takes a Hi3D mesh and turns it into an interactive world. The only honest overlap is at the top of a project: both can inform a decision about what a scene should look like before anyone commits to building it. One shows you the shot; the other gives you one object from it.

Cost of ownership, which is where the choice is really made

Hi3D v1.0 has a price you can calculate today: the cost of a 96 GB GPU for the duration of the refinement, plus the engineering time to assemble the inference path, plus whatever the licence requires once your product crosses its threshold. That is knowable, budgetable and repeatable. The trade is that you own the pipeline and its failure modes.

Genie 4 has no calculable cost. There is no API price, no subscription price specific to the model, and no usage meter published. The only number attached to access is the AI Ultra subscription that currently reaches Genie 3 through Project Genie, and Google has not said whether Genie 4 will be included at the same tier or priced separately. A team cannot put that in a budget, which is a legitimate reason to wait even if the demos are exactly what the project needs.

Neither model is in OrcaRouter's catalogue, and it is worth saying why rather than leaving a gap: we do not host a 3D generation or refinement model today, and Genie 4 has no public endpoint to route. What the catalogue does carry is the hosted video-generation tier — models such as Kling v3 that turn a prompt into a clip — and those are prompt-to-video tools, not interactive world models. Treating them as a Genie 4 substitute would be a category error, so we will not. What we can offer is the surrounding stack: one API across 200+ models with provider list prices passed through at 0% markup, automatic failover across providers, and a routing DSL for composing several models into one call. If a pipeline needs a captioner, a planner and a video model, that composition is the part we can actually help with.

The decision, then, is not close. If the deliverable is a mesh that a printer, a game engine or a review pipeline will consume, Genie 4 is not in the running and Hi3D v1.0 is the only one of the two that produces the thing you need — provided you have the GPU and can live without texture. If the deliverable is a judgment about how a place moves and sounds, Hi3D v1.0 is not in the running and Genie 4 is the only one that shows you, once the waitlist opens. Buy the artifact, not the demo.