Hunyuan3D Review: Open Weights, License and Hardware

Which “Hunyuan3D” are we talking about?

The name covers several things, and mixing them up leads to wrong license assumptions. Open repositories from Tencent include Hunyuan3D-2.1 (released June 13, 2025), Hunyuan3D-2, Hunyuan3D-Omni (released September 25, 2025), plus Part, 2mini, 2mv and 1. A 3.0 release was announced on September 17, 2025 as a hosted product, and Tencent Cloud’s billing page mentions 3.0 and 3.1 as selectable model versions on the hosted API. We found no 3.x open weights in Tencent’s GitHub organization or Hugging Face listing, and a 2.5 appears only as a technical report. As of November 2026, the latest downloadable release we could confirm is 2.1, plus Omni. Hosted 3.0 and 3.1 feature specs are only supported by press and a billing page, not by primary documentation.

License: what the 2.1 text says

The 2.1 repository ships the “Tencent Hunyuan 3D 2.1 Community License Agreement” (June 13, 2025). The 2.1 README has no license section, so the LICENSE file is the source. Key points, from the file as relayed:

• Territory: “the worldwide territory, excluding the territory of the European Union, United Kingdom and South Korea.” The header reads “DOES NOT APPLY IN THE EUROPEAN UNION, UNITED KINGDOM AND SOUTH KOREA,” and use outside the Territory is “unlicensed and unauthorized under this Agreement.”
• Scale: if your products or services had more than 1 million monthly active users in the calendar month before that version’s release date (June 13, 2025 for 2.1), you must request a license from Tencent, at its sole discretion, with no rights until granted. This is a paraphrase from the fetch.
• Outputs: section 6(d): “Tencent claims no rights in Outputs You generate.” Section 5(b) prohibits using the works or output to improve any other AI model, other than Hunyuan3D 2.1 or its derivatives.
• Commercial use: the grant does not use the word “commercial” per our fetch; commercial use is governed by the Territory and monthly-user terms.

The 2.0 license has the same structure. Omni’s Hugging Face tag is tencent-hunyuan-community, but its terms text was not shown there. We did not read the exact section 4 wording, only a paraphrase.

What it means for a shipped game: on the license text as relayed, a developer outside the excluded regions with fewer than 1 million monthly users appears to be licensed to use the output, and Tencent claims no rights in it. A developer in Germany, Britain or Korea cannot rely on this license at all. The ban on training another model on outputs matters if you build a style-tuned model from the assets. See Hunyuan3D and TRELLIS license terms. This is not legal advice, and the hosted API has its own terms that we did not read.

Hardware and running it locally

Tencent’s README for 2.1 says “It takes 10 GB VRAM for shape generation, 21GB for texture generation and 29GB for shape and texture generation in total,” tested on Python 3.10 and PyTorch 2.5.1 with CUDA 12.4. Hunyuan3D-2 needed 6 GB for shape and 16 GB total. Omni needs 10 GB. Setup involves pip install -r requirements.txt, building a custom rasterizer and mesh painter, downloading a Real-ESRGAN checkpoint, and running the Gradio app, which has a --low_vram_mode flag. ComfyUI integrations for 2.1 appear in the README as community projects. What this means for a job: the README’s separate 10 GB (shape) and 21 GB (texture) figures against 29 GB combined suggest, as our inference only, that the stages are what a smaller card would have to run one at a time; the README does not say so. A 40-prop kit on a 24 GB card is therefore a question the README does not answer, and the hosted API below is the documented fallback. Whether low-VRAM mode on a 12 or 16 GB card produces usable results is unmeasured; the planned local-run test targets it.

Output

The 2.1 README claims “production-ready PBR Material,” and the pipeline is image to shape, then a separate paint pipeline for texture. File formats are not specified in the 2.1 README; the 2.0 code saves to GLB or OBJ through trimesh, and the hosted API returns GLB. We have no output-quality evidence of our own and no source we trust for independent numbers. What survives the next step in Blender or an engine is covered in importing AI models into Blender.

Hosted API alternative

If your GPU is too small, Tencent’s HY 3D Global API on Tencent Cloud has a one-time 200-credit free allowance valid for a year, prepaid packs (1,000 credits for $15, up to 100,000 for $1,350), and postpaid at $0.02 a credit. Professional-mode generation costs 25 credits for normal textured output (so 8 free generations), 30 for LowPoly and 15 for geometry only, with add-ons of +10 for PBR, multi-view images or face count. At the $15 pack rate, a textured PBR generation (35 credits) is about $0.53; postpaid it is $0.70. Post-processing: part generation 30, smart topology 50, texture edit 30, UV unfold 10, auto-rigging 10, format conversion 5. Default concurrency is 3. A 40-prop kit at one attempt each, using the normal textured generation (25 credits) plus PBR (+10), is 1,400 credits: about $21 at the $15-per-1,000 pack rate or $28 postpaid at $0.02, against 200 free credits, which cover five such generations. These are our planning assumptions from the billing page, not measurements. The consumer site 3d.hunyuan.tencent.com was unreadable, so its pricing and free terms are unverified. 3D AI Studio also lists a Hunyuan model; see its review.

Alternatives

TRELLIS.2 is MIT-licensed but needs 24 GB; the Hunyuan3D vs TRELLIS.2 comparison covers the choice, and open-source generators you can run yourself lists the field. Our methodology describes the planned checks; none has been run.

More generators are listed in the AI 3D generator reviews hub.