ai-info

Self-hosting

Hi3D V1.0: The 96GB H20 Open-Source 3D Refiner

·

Verdict: Hi3D V1.0 is a genuinely capable geometry refinement model open-sourced under a "production-validated" banner — and a case study in reading the license before tweeting "open source." The Hi3D Open License Agreement 1.0 is custom, not OSI: free for nearly everyone, but cross 200,000 monthly active users or $1M monthly revenue and you owe the licensor (OCEANO HK Holding) a separate written license application within 60 days. The steeper wall is hardware: the 1536³ resolution ladder has only been validated on a 96GB H20, there's no CPU path, flash-attn has no fallback, and the coarse stage isn't even Hi3D's — it's Tencent Hunyuan3D-2 under its own separate license. If you're a 3D printing or e-commerce shop below the threshold with access to serious GPUs, this is real production-grade geometry for free. If you were hoping for a 4090 workflow or a clean Apache baseline, neither is on the menu.

Last checked 2026-10-31. License text and hardware requirements verified against the official release; scores below are the vendor's unless marked otherwise.

What actually got released

On October 23, 2026, the Hi3D team (Chinese company, licensor of record OCEANO HK HOLDING LIMITED) open-sourced Hi3D V1.0 — the refinement half of the pipeline formerly known as Sparc3D (the Python package still imports sparc3d for compatibility). The model takes a reference image and a coarse mesh, and progressively refines geometry up to an effective sparse extraction-grid resolution of 1536³, via three passes: 512³ (one pass), 1024³ (two), 1536³ (all three). Training reportedly scaled from ~500K to ~2 million curated 3D assets.

What's in the box: model artifacts, inference code (GitHub), docs. What is not in the box, per the README: training code, and the paper's evaluation harness. And the coarse stage that feeds it — the part that turns your image into the initial mesh — is Tencent's Hunyuan3D-2, governed by Tencent's licenses, not Hi3D's. You're assembling a two-license sandwich before you've rendered your first cube.

The license, minus the marketing

The press release calls it "the first production-validated open-source AI 3D model." The LICENSE file tells a more precise story:

  • Grant is broad. Use, modify, fine-tune, distribute, sublicense, sell, host — APIs and SaaS explicitly included. Outputs (meshes, exported files) are yours; standalone outputs don't infect your finished work.
  • The tollbooth is scale, not usage type. Trigger 200,000 MAU in a month or $1M monthly revenue across related products, and you must notify the licensor and apply for a separate written license within 60 days — with disclosure of products, MAU methodology, deployment regions, security measures. Splitting a service across brands "to evade the Scale Threshold" is explicitly defined as aggregation-bait (Section 1.9).
  • Attribution mechanics. Distribute the materials or derivatives and you carry the full agreement text, a NOTICE file, and marked modifications downstream.
  • Training other models is permitted — to the extent the licensor owns the rights — but outputs that passed through third-party models (your Hunyuan coarse meshes included) don't inherit this permission.
  • Patent grant terminates the day you sue the licensor or a contributor over the materials. Standard defensive clause, worth knowing before you build a business on it.

For calibration: this is the same species as OXR's license we dissected yesterday — "open" with a meter. Hi3D's threshold ($1M monthly revenue / 200K MAU) is more generous than OXR's ($2M total funding or revenue, plus a ban on competing OCR APIs). But neither is Apache-2.0, and "open source" press releases won't hold up as legal advice.

The hardware bill is the real bill

The README's requirements section is unusually honest, which we appreciate: Linux, Python 3.10, NVIDIA compute capability 8.0 or newer, CUDA 12.0–12.5, 96GB H20 as the measured and validated configuration — and the 1536³ ladder "has only been validated on that GPU; peak memory and feasibility on smaller cards have not been established." There is no CPU inference path. flash-attn kernels are mandatory with no fallback. Offloading via Accelerate reduces residency but, per the docs, "does not establish a minimum supported GPU memory size."

Translation for planning purposes: this is a rented-datacenter-GPU workload, not a workstation tool. A 24GB card (the entry floor for most "local LLM" guides, ours included) is not in the conversation until someone publishes a validated smaller-card path. Budget accordingly, and treat any smaller-GPU deployment as your own R&D project, not a supported configuration.

What it's good at, and what to check yourself

The vendor's reported evaluation claims the highest multi-view fidelity and detail-richness scores, plus a 90% slicing success rate (the 3D-printing metric that actually matters — meshes that survive the slicer). Those are vendor numbers; the evaluation harness isn't in the release, so independent reproduction isn't trivially possible. The geometry itself — layered relief, scales, teeth, concentric grooves — is what the comparison images showcase, and the technical report documents the sparse-grid approach honestly (nominal structural grids at 64³/128³/192³ rather than a magic dense 1536³ tensor; read the fine print before quoting "1536³" in your own marketing).

Two operational quirks worth knowing: a request that can't reach target resolution raises an error rather than silently returning lower resolution (good — no mystery-quality meshes), and GPU sparse convolution is not bitwise deterministic, so compare meshes geometrically, not by hash.

Who should bother

3D printing and e-commerce product pipelines below the scale threshold, with H20/B200-class access: strong fit, and free beats every commercial API's per-mesh pricing at volume.

VR/AR and animation studios: evaluate, but note the refinement model needs a coarse mesh from somewhere — the Hunyuan stage is the default, and its license terms are a separate homework assignment.

Anyone at $1M monthly revenue or 200K MAU: talk to the licensor before shipping, not after. The 60-day clock starts at the trigger, and the application requires more disclosure than a typical sales call.

Everyone hoping "open source" means Apache/MIT: it doesn't here. The custom agreement is reasonable as these things go — but it's a contract with a named counterparty, and you should read it as one.

Sources: Hi3D Open License Agreement Version 1.0 (full text via Hugging Face hi3d-official/Hi3Dv1.0 LICENSE, dated 2026-10-22 — scale threshold §5, grant §2, aggregation §1.9, patent §3.1); Hi3D v1.0 README (hardware requirements, three-pass pipeline, no-training-code notice, Hunyuan3D-2 third-party stage, determinism caveats — pulled 2026-10-31); HF model API (license: other, hi3d-open-license-agreement-1.0, 140 likes); PRNewswire release 2026-10-23 and Chinese coverage (500K→2M assets, 90% slicing success rate, evaluation claims — transcribed vendor claims, not independently verified).

Related reading

OXR-1.0: The 1.3B Document Parser Topping Three Benchmarks — Deployment Facts and License Strings Before You Ship ItOXR-1.0 ranks #1 on olmOCR-Bench, OmniDocBench and DocParsingBench with an RSI post-training loop — but the license isn't plain Apache-2.0 ($2M threshold, attribution duty, no competing OCR API) and the official VRAM floor is 24GB, not the RTX 3060 you may have read. Deployment facts, license strings, and honest limits.MOSS-Image Is the Open-Weight Image Model You Can Actually Fine-Tune at Home — but Budget for the 9B Text Encoder, Not the 5B DiTMOSS-Image (OpenMOSS, Apache-2.0) topped Hugging Face trending in four days. The local deployment math: 5B DiT plus a 9B Qwen3.5 text encoder, LoRA fine-tuning from 6GB VRAM, day-0 ComfyUI/SGLang/vLLM support, and the community quantization ecosystem. What GPU you actually need.