Hi3D V1.0 Goes Open Source: Free AI 3D Models — With a $1M License Catch

On October 23, 2026, Beijing-based Hi3D (Mathmagical) announced via PRNewswire that Hi3D V1.0 — the AI 3D model previously known as Sparc3D — is now open source. The inference code is public on GitHub, the weights are on Hugging Face, and the same day the company announced it is cutting credit consumption on its hosted platform by over 50%.

It is the kind of announcement that reads like pure generosity — a "production-validated" 3D model, free, with a technical report. But the details matter here more than usual, because this release is a geometry refiner, not a 3D generator, and its license is custom — not MIT, not Apache. Here is what you actually get, what it costs to use, and where the strings are.

Quick answer

• What it is: Hi3D V1.0, an image-to-3D geometry refinement model open-sourced on October 23, 2026. Inference code, weights and documentation are free to download; training code is not included.
• What it does: It takes a coarse mesh plus a reference image and refines it into high-resolution geometry at up to 1536³. It does not start from a bare photo — you must supply (or generate) the rough shape first.
• The price: $0 for the model itself, but it ships under the custom Hi3D Open License Agreement v1.0. Commercial use is permitted until a product passes 200,000 monthly active users or $1M in monthly revenue — then you must apply for a separate license.
• The hosted alternative: The Hi3D web platform charges subscriptions — Pro at $9.9/month for new users ($19.9 on renewal) with 1,000 monthly credits — and the free tier includes 100+ credits. Generation is full image-to-3D there, textures included.

What actually got released

The GitHub repository hi3d-official/Hi3Dv1.0 (created October 16, 2026; 196 stars and 16 forks at the time of writing) contains inference code, eight model artifact files on Hugging Face, documentation, Docker setup, and demo scripts. The README — which is refreshingly candid — says explicitly: this repository provides inference code only. Training code and the paper's evaluation harness are not included. That means you can run the model and inspect its outputs, but you cannot retrain it or reproduce its benchmark tables from the repository alone.

Two conditioning encoders — facebook/dinov2-large and openai/clip-vit-large-patch14 — are loaded separately from Hugging Face on first use and are not part of the model package. The Python package keeps the old Sparc3D names (sparc3d, Sparc3DPipeline) for compatibility.

What the model does — and what it deliberately doesn't

Hi3D V1.0 is a geometry refinement model. Give it a coarse mesh and a reference image of the same object, and it progressively samples sparse latents, decodes them with a self-pruning sparse VAE, and extracts a refined mesh — in three passes labeled 512³, 1024³ and 1536³. Those numbers describe nominal sparse coordinate and extraction grids, not dense 1536³ tensors or fixed triangle counts — the README says so directly, which is worth noting because press releases rarely qualify resolution claims.

What comes out is geometry only: no textures, no materials, no vertex colors. The output is recentered and normalized to a unit maximum radius, so world coordinates and absolute scale are not preserved. And the coarse input matters: the docs warn that severe structural errors in the coarse mesh can persist, and geometry hidden from the reference image is a plausible completion, not a reconstruction.

For the full end-to-end demo — a single image in, a mesh out — the repository includes an optional adapter for Tencent's Hunyuan3D-2 mini as the coarse generator. That code and those weights are third-party and carry their own license; they are not part of what Hi3D released. There is also a live demo on Hugging Face Spaces if you want to see it without touching a GPU.

The license: the most important part of this story

Here is where "open source" needs its quotation marks. Hi3D V1.0 is governed by the Hi3D Open License Agreement Version 1.0 — a custom agreement, not an OSI-approved open-source license. GitHub's own metadata lists the repo license as "Other (NOASSERTION)". The original Apache-2.0 and MIT terms remain in force for the original and third-party portions they cover, but Hi3D's own portions — the second-stage implementation, the docs, and the eight model artifact files — fall under its agreement. The key terms:

• Commercial use is permitted — until scale. If a related product or service exceeds 200,000 monthly active users or USD 1,000,000 in monthly revenue, you must notify the licensor and apply for a separate written license within 60 days.
• Commercial integrations must display "Powered by Hi3D 1.0" branding, linking to the official website where supported.
• Standalone outputs get different treatment: obtaining, editing, displaying, selling or distributing an output that contains no Hi3D materials or model derivatives does not by itself trigger the branding or scale-license obligations. In practice: ship a mesh, you're fine; ship a product built on the model past the thresholds, you owe Hi3D a licensing conversation.

For license questions the repo gives a contact: [email protected]. Compare this to permissive releases like Meta's Llama community license (which also has usage thresholds) — the pattern is familiar, but call it what it is: open weights under a custom commercial license, not open source in the OSI sense.

Pricing: the hosted platform and the 50% credit cut

Hi3D's announcement paired the open-source release with a price move on its hosted platform: a reduction of over 50% in credit consumption, "lowering the barriers to 3D creation." The announcement does not publish new per-generation credit numbers, so the concrete math below comes from the rates Hi3D's own Blender plugin guide published for its API credit packages (v1.5/v2.0 era) — treat these as pre-cut baselines that should now fall by half or more:

• Geometry, single image or multi-view, 512³: 5 credits
• Geometry at 1536³ (v1.5 or v2.0): 40 credits
• Geometry at 1536³ Pro (v2.0): 60 credits
• Texture generation (v2.0): 20 credits, or 30 with PBR materials

Subscription plans, per product listings on Slashdot, TopBusinessSoftware and SourceForge (all consistent as of October 2026):

Hi3D's official FAQ confirms the headline terms: new users get 100+ credits on the Free plan, subscription credits refresh each billing cycle, and — importantly for a credit-burn product — failed generations are automatically refunded, with 3 to 15 free retries depending on plan. "You own the models you create," the FAQ says; paid plans allow private and commercial use, while free-plan models are subject to the free asset license and require attribution.

The benchmarks: vendor numbers you cannot check

The announcement claims Hi3D V1.0 achieved "the highest multi-view fidelity and detail richness scores in the reported evaluation" alongside a 90% slicing success rate, with training expanded from roughly 500,000 to two million curated 3D assets. All of this is company-reported, and — per the README — the evaluation harness, benchmark splits and comparison meshes were not released. Treat the benchmark claim as marketing until independent testing exists. To its credit, the project documents its limits better than most: the 1536³ ladder "has only been validated" on the measured GPU, and "peak memory and feasibility on smaller cards have not been established."

The hardware bar for self-hosting

This is not a model you run on a laptop. Requirements: Linux, Python 3.10, an NVIDIA GPU with compute capability 8.0 or newer, and CUDA 12.0–12.5. There is no CPU inference path, flash-attn has no fallback, and the measured stack used a 96 GB NVIDIA H20. If your GPU has less memory, you are in uncharted territory — the project's own docs say so. For most creators, the hosted platform or the free HF demo will be the practical route; the open weights matter most to labs and companies that want private deployment without vendor lock-in.

Three ways to use Hi3D, compared

Pros and cons

Pros: genuinely downloadable weights and code, not a marketing "open"; unusually honest documentation that states its own limits; commercial use permitted under a clear (if custom) license; >50% credit-consumption cut on the hosted platform; no-texture geometry focus is exactly what 3D-printing and game pipelines need.

Cons: not OSI open source — custom license with a $1M/200K-user trigger and mandatory "Powered by Hi3D 1.0" branding; geometry-only output (no textures); requires a coarse mesh, so it is not true image-to-3D; GPU-only with a steep, barely-tested hardware bar; vendor benchmark claims cannot be independently reproduced; per-generation credit numbers predate the announced 50% cut.

FAQ

Is Hi3D V1.0 really free?

Yes — the inference code and weights cost nothing to download and run. But "free" comes with the Hi3D Open License Agreement v1.0, which requires "Powered by Hi3D 1.0" branding on commercial integrations and a separate license once your product passes 200,000 monthly active users or $1M in monthly revenue. Standalone meshes you produce are not affected.

Can I use Hi3D in a commercial product?

Under the license, yes — commercial use is explicitly permitted. The obligations kick in at scale (the 200K-user / $1M-revenue thresholds) and in the form of the branding requirement. If your use case involves the hosted platform rather than the open model, paid plans allow private and commercial use; free-plan assets require attribution.

What hardware do I need to run Hi3D V1.0 myself?

Linux, Python 3.10, an NVIDIA GPU with compute capability 8.0 or newer, and CUDA 12.0–12.5. There is no CPU path. The project validated the 1536³ ladder only on a 96 GB NVIDIA H20, so anything smaller is unverified territory. For a quick look without hardware, use the Hugging Face demo.

Does it generate a full 3D model from a single photo?

The open model does not: it refines a coarse mesh you supply, guided by a reference image. Since image quality drives the result, it pays to start with a strong source image — our roundup of the best free AI image generators and our FLUX 3 Max review cover the tools that can produce one. The end-to-end demo pairs it with Tencent's Hunyuan3D-2 mini for the coarse stage (separate license). The hosted platform does offer full image-to-3D from a photo, sketch or text prompt, with textures and a 3D-print workflow including STL/OBJ/3MF/GLB/USDZ export and one-click sending to Bambu Studio, OrcaSlicer or Creality Print.

How much does a generation cost on the platform?

After the announced >50% credit cut, exact numbers are not yet published. Before the cut, Hi3D's own guide listed 5 credits for a 512³ generation and 40 for 1536³ on the API credit packages. On the Pro plan ($9.9/mo intro, 1,000 credits), that arithmetic put a top-resolution geometry pass at roughly 25 generations a month pre-cut — the new numbers should roughly double that.

Key takeaways

• Hi3D open-sourced Hi3D V1.0 (formerly Sparc3D) on October 23, 2026: inference code, weights and docs are free — training code and the evaluation harness are not.
• The model refines a coarse mesh into high-resolution geometry at up to 1536³; it is geometry-only, needs a reference image, and does not generate textures or start from a bare photo.
• The license is custom, not MIT/Apache: commercial use is allowed, but the "Powered by Hi3D 1.0" branding and the 200,000-user / $1M-revenue relicensing trigger apply.
• The hosted platform cut credit consumption by over 50%; plans run from a free 100+ credits to Ultra at $129.9/month, with Pro at $9.9/month intro ($19.9 on renewal).
• Self-hosting demands an NVIDIA GPU with CUDA 12.0–12.5 and no CPU path — validated only on a 96 GB H20 — so most creators should start with the free demo or the hosted free tier.

October 24, 2026 — Open weights are only as useful as their license and their docs. Hi3D deserves credit for publishing both honestly — its README states every limitation that the press release skipped. For 3D creators, the real news may be less the open model than the platform's 50% credit cut: image-to-3D at half the price, with the weights as a safety net against lock-in. Just read the license before you build a business on it.