Hi3D v1.0 vs Hy-Nomas: Both Say Open Source, Only One Fits Your GPU

Hi3D v1.0 and Hy-Nomas were published two days apart in October 2026, both from Chinese labs, both with downloadable weights, and both described by their publishers as open source. They are not the same kind of open. Hy-Nomas, Tencent Hunyuan's interactive world model published on 21 October 2026, ships under Apache 2.0 and documents a path onto a single RTX 4090. Hi3D v1.0, the Sparc3D line open-sourced on 23 October 2026, ships under a custom agreement with a 200,000-monthly-user trigger and was validated on a 96 GB H20. One of those differences decides whether you can deploy at all. The other decides whether you can deploy commercially without a phone call.

Neither model is a substitute for the other, and the licence and hardware comparison is more useful than a quality comparison, because neither publisher has released evidence that would support a quality comparison. So start with the doors, not the pictures.

Two licences, two different doors

• Hy-Nomas — Apache 2.0. No user cap, no revenue trigger, no attribution requirement beyond the usual notice obligations. This is a licence a legal team can approve without a conversation.

• Hi3D v1.0 — the Hi3D Open License Agreement 1.0, which is not an OSI-approved licence. Commercial use is permitted. But if a related product or service exceeds 200,000 monthly active users or USD 1,000,000 in monthly revenue, the licensee must notify Hi3D and apply for a separate written licence within 60 days. Commercial integrations must also display "Powered by Hi3D 1.0," linking to the official site where the medium supports it. Standalone outputs that contain no Hi3D materials or derivatives do not by themselves trigger the branding or scale obligations.

The Hi3D terms are not hostile. Below the threshold they are permissive, and the output carve-out is unusually sensible for a tool whose whole job is to produce a file. But they introduce an obligation that switches on with success, which is a diligence item for any funded company and a genuine unknown for a marketplace or a consumer app. Apache 2.0 has no such switch. If your deployment could plausibly cross two hundred thousand monthly users, that difference is worth more than any benchmark either vendor has published.

The hardware bar is the real moat

• Hy-Nomas — eight GPUs by default. A single RTX 4090 is supported with INT8 or FP8 quantization plus CPU offload. Full checkpoints are around 196 GB; the low-memory path is documented at roughly 38 GB. That is still a serious machine, but it is a machine a small studio or a well-equipped individual can own.

• Hi3D v1.0 — a 96 GB NVIDIA H20 in the validated configuration. Linux only, compute capability 8.0 or newer, CUDA 12.0 to 12.5, no CPU inference path, batch size one, and flash-attn/ with no fallback attention implementation. Sequential CPU offload reduces GPU residency but the README explicitly says it does not establish a minimum memory requirement.

This is the widest practical gap between the two releases. A 4090 owner can run Hy-Nomas at reduced precision and accept the quality cost. A 4090 owner cannot run the configuration Hi3D validated: the 1536³ path has only been tested on the H20, and peak memory on smaller cards is unstated. The 512³ one-pass setting may fit, but the vendor has not claimed it and nobody has published a memory table. If you are budgeting, Hy-Nomas has a documented cheap path and Hi3D does not.

What comes out

• Hi3D v1.0 — a mesh. Geometry only, no textures, no materials, no vertex colours, recentered and normalized to a unit maximum radius. Scale and world coordinates are discarded, so the file has shape but no physical size.

• Hy-Nomas — video. 720×1280 at 24 frames per second, eight denoising steps in the distilled configuration. There is no mesh, no depth buffer and no coordinate system in the output. The model generates appearance over time, not geometry.

Neither one hands you a textured asset, and the failures are inverse. Hi3D gives you shape without surface: printable, but grey. Hy-Nomas gives you surface without shape: convincing to watch, useless to manufacture. If your downstream tool is a slicer, only one of these outputs can feed it. If your downstream tool is a video encoder, only the other can.

What goes in

• Hi3D v1.0 — a reference image plus a coarse mesh of the same object. The repo includes an optional adapter that uses Tencent's Hunyuan3D-2 mini to produce that coarse mesh; Hunyuan is third-party and not a dependency of the core library, but without a coarse stage there is no input.

• Hy-Nomas — multi-turn free-form text plus an explicit camera trajectory, supplied as camera-to-world matrices and intrinsics through cross-attention. There is no image input. The model is steered by language and by where you point the virtual camera.

This is a cleaner distinction than the output comparison, because it tells you what kind of authoring each tool expects. Hi3D expects an object that already exists in rough form and a photograph of it. Hy-Nomas expects a description and a path through a scene. One is an asset tool; the other is a camera.

What the repos actually ship

• Hy-Nomas — a 54-block dual-stream MMDiT backbone with hidden size 2,048 and 16 heads, a Qwen2.5-VL-7B-Instruct text encoder, Cinematic and Robotics LoRAs, a LoRA training recipe, and Hy-Nomas-Bench: 280 cases with about 220 reference camera trajectories. The benchmark ships without published scores, and the technical report is marked as coming soon. The repository opened with a single initial commit.

• Hi3D v1.0 — inference code, eight weight files, Docker constraints, demos and documentation. Training code, the evaluation harness, benchmark splits and comparison meshes are all absent, and no parameter count is published anywhere in the repo, model card or release.

Neither repository lets you reproduce its headline numbers, which is the honest common ground. The difference in kind is that Hy-Nomas ships a benchmark set and a fine-tuning recipe, so you can at least build your own evaluation and adapt the model. Hi3D ships a model you can run and nothing you can adapt. If your team intends to fine-tune, that gap is decisive.

Deploying either one, honestly

If you need a world to explore, Hy-Nomas is the only one of the two that makes one, and its low-memory path is the reason it is deployable outside a cluster. If you need an object, Hi3D v1.0 is the only one that produces geometry, and the H20 requirement is the reason it is not a laptop tool.

What neither does is complete a pipeline. Hi3D needs a coarse mesh and gives you no texture. Hy-Nomas needs a scene description and gives you no geometry. A studio that wants both ends would run Hy-Nomas for context and Hi3D for hero props, with a coarse-mesh generator in between and a texture stage after — four models, three of them not in this comparison.

OrcaRouter does not host either model, and saying otherwise would be false. What the platform does provide is the non-3D work around them: prompt expansion before Hy-Nomas, reference-image captioning before Hi3D, and vision checks on rendered output afterward. That runs on 0% markup on provider list price, so a vendor price change is live on our side the same day, and the routing DSL lets you compose those steps into a single call rather than stitching four SDKs together. The 3D weights still run on your hardware.

The thing to watch

Hy-Nomas is the more conventionally open release, and that is likely to matter more over time than its current lack of published scores: Apache 2.0 plus a LoRA recipe is a platform, while a custom licence plus inference-only code is a product. Hi3D's threshold is generous enough that most users will never touch it, but the moment one integrator crosses 200,000 monthly users, the terms change and the "open source" label becomes materially misleading.

For now the decision is simple. If your constraint is a single GPU and a clean licence, Hy-Nomas is the one you can actually run. If your constraint is a mesh file at the end, Hi3D v1.0 is the only one of the two that produces one, and you should budget for the H20 before you budget for anything else.