Buying guide · Generators
Open-Source AI 3D Generators You Can Run Yourself
Open-weight AI 3D generators you can self-host: Hunyuan3D 2.1, Omni, TRELLIS and TRELLIS.2, compared on license terms, documented VRAM, OS support and outputs. Nov 2026.
If you want to run an AI 3D generator on your own machine, the two documented choices are Microsoft’s TRELLIS.2 (MIT license, at least 24 GB of GPU memory, Linux-tested) and Tencent’s Hunyuan3D 2.1 (staged memory use from 10 GB, but a territory-limited license). Which is right depends mostly on your GPU and where you sell. No affiliate programs apply to this page, and we have not installed or run either model.
How the picks were made
The criteria are the ones the repositories document: license for code, weights and outputs; stated VRAM; operating system; inputs; outputs; and whether integrations are official. “Open source” here is loose. Tencent’s license is a custom community license, not an OSI-style open-source license, so the more accurate term is open weights. See the methodology for our evidence approach.
| Model | License | Documented VRAM | OS / tested stack | Inputs and outputs | Best for |
|---|---|---|---|---|---|
| TRELLIS.2 (4B) | MIT; nvdiffrast and nvdiffrec limited to non-commercial use | At least 24 GB (A100, H100 verified) | Linux only tested; CUDA 12.4, PyTorch 2.6.0 | Single image to PBR GLB; shape-conditioned texturing | Permissive model license on a 24 GB card; check dependency limits |
| Hunyuan3D 2.1 | Tencent Community License; excludes EU, UK, South Korea; over 1M monthly users needs a license | 10 GB shape, 21 GB texture, 29 GB both | README: macOS, Windows, Linux; tested with Python 3.10, PyTorch 2.5.1+cu124; --low_vram_mode flag | Image to shape, then PBR paint; model tags also list text-to-3D | Smaller GPUs, outside excluded regions |
| Hunyuan3D 2.0 | Community license (2.0 terms not read for this page) | 6 GB shape, 16 GB both | Not specified in README | Image to 3D; mesh saved as glb/obj | The lowest documented memory floor |
| Hunyuan3D-Omni (3.3B) | Tag tencent-hunyuan-community; license text on Hugging Face matches the 2.1 Community License apart from date, version name, repository link and a former-entity footnote | 10 GB | Not specified | Image plus optional point cloud, voxel, box or pose control | Controlled generation |
| TRELLIS (v1) | MIT for models and most code; the diffoctreerast dependency is research-only | At least 16 GB | Linux only tested | Text or image prompts; Gaussians, radiance fields, meshes; sample GLB | Text prompts on a 16 GB card |
TRELLIS.2: permissive license, high floor
TRELLIS.2’s README states: “An NVIDIA GPU with at least 24GB of memory is necessary,” verified on A100 and H100 and “currently tested only on Linux.” The code and model are MIT-licensed. Its model card is candid about weaknesses: raw meshes “may occasionally contain small holes or minor topological discontinuities,” and the base model “has not been aligned with human preferences.” Plan for hole and floater repair. Official ComfyUI support is not stated. See the TRELLIS.2 review.
Hunyuan3D 2.1: staged memory, restrictive license
Hunyuan3D 2.1 documents separate memory figures for shape (10 GB), texture (21 GB) and both (29 GB), plus a --low_vram_mode flag whose memory saving the README does not quantify. The shape stage alone fits well under TRELLIS.2’s 24 GB figure, but the full textured pipeline (29 GB) asks for more. The cost is the license: the grant covers “the worldwide territory, excluding the territory of the European Union, United Kingdom and South Korea,” with a license request required above one million monthly active users, and Section 5(b) bars using outputs to improve other AI models. Section 6(d) says Tencent claims no rights in your outputs. The 2.1 README lists a ComfyUI node only as a community contribution. The newest Hunyuan releases, 3.0 and 3.1, appear on hosted services (3D AI Studio lists 3.1); we did not find open 3.x weights. See the Hunyuan3D review, the license notes and the head-to-head Hunyuan3D vs TRELLIS.2.
What each pick means for the destination
TRELLIS.2’s documented output is sample.glb, described as a PBR-ready asset exported opaque by default, and glTF 2.0 is the format Godot’s manual recommends; Unity’s manual recommends FBX and does not mention glTF, so a Unity project would need a conversion step that neither repo documents. Hunyuan3D 2.1’s README does not specify output formats (2.0’s code returns a mesh that can be saved as glb or obj). Neither README documents rigging, LODs or print checks, and TRELLIS.2’s own model card says raw meshes may contain holes. For a shipped game, TRELLIS.2’s MIT grant is the simpler basis; Hunyuan3D 2.1’s license excludes the EU, UK and South Korea outright, which matters if you or your players are there, and bars using outputs to improve other AI models. Not legal advice; the licensing note covers the wording.
Downsides to weigh
Self-hosting removes per-run fees but not cost: you supply the GPU, a Linux setup for TRELLIS (the only tested OS in its README), a custom rasterizer build for Hunyuan3D 2.1, and your own cleanup. Neither TRELLIS README mentions ComfyUI, and Hunyuan3D 2.1’s README lists its ComfyUI node only as a community contribution.
What “documented VRAM” does and does not mean
These are vendor-stated requirements. They are not measured peaks on a consumer card, and the TRELLIS READMEs name Linux testing and datacenter GPUs (A100, H100; A6000 for v1). Hunyuan3D 2.1’s README says it “supports Macos, Windows, Linux” and that it was tested with Python 3.10 and PyTorch 2.5.1+cu124, but names no GPU model. Whether a 12 or 16 GB GeForce card completes a full texturing run in reasonable time is exactly what a local-run test would answer, and ours is planned but not done: open-source local run.
Picks
Pick TRELLIS.2 if you have a 24 GB NVIDIA GPU, run Linux, and want MIT terms. Consider Hunyuan3D 2.1 if you are outside the excluded territories, can accept the user-count terms, and want a staged pipeline; on documented numbers it fits smaller cards only for the shape stage (10 GB) or the texture stage (21 GB) run separately, and Hunyuan3D 2.0 documents 16 GB for both under a license whose terms we did not read. The basis is the READMEs and license files, not any run. If you have neither the hardware nor the appetite for setup, a hosted tool is simpler; see best AI 3D model generators and free AI 3D generators.
What evidence would change this
A measured install and local run (time, peak memory, errors, outputs) on a consumer GPU, or an official statement on sub-24 GB operation of TRELLIS.2, would change the picks. Other open models exist that we have not reviewed. Back to best-of guides.