Microsoft TRELLIS.2 review: the best free, open AI 3D modeler you can self-host
Answer in brief: Microsoft TRELLIS.2 is the best AI 3D modeler if you have the hardware and want zero per-model fees. It is an open 4-billion-parameter image-to-3D model that generates fully textured assets with PBR materials (base color, roughness, metallic and opacity) at 512³ to 1536³. Microsoft reports about 3 seconds per model at 512³ on an H100. The code is MIT-licensed, the weights are on Hugging Face and the training code is public. The cost is setup: Linux, an NVIDIA GPU with at least 24 GB, image input only, and heavy meshes to decimate. This is independent research from Microsoft's repository; we did not run it ourselves.
Verdict: maximum control, zero per-model cost. Choose TRELLIS.2 for research, private pipelines or very high volume on your own GPUs.
What TRELLIS.2 does
TRELLIS.2 uses a 'field-free' sparse voxel structure Microsoft calls O-Voxel. It reconstructs and generates assets with complex topologies, sharp features and full PBR materials. It handles open surfaces such as clothing and leaves, non-manifold geometry and internal enclosed structures. A separate pipeline generates PBR textures for a 3D shape you already have. Exports are GLB files.
| Resolution | Total time on H100 (Microsoft) | Shape + material |
|---|---|---|
| 512³ | About 3 s | 2 s + 1 s |
| 1024³ | About 17 s | 10 s + 7 s |
| 1536³ | About 60 s | 35 s + 25 s |
Who TRELLIS.2 fits best
- ML engineers and researchers who want to fine-tune or train on their own data.
- Studios with GPU servers generating at volumes where per-call fees add up.
- Privacy-sensitive teams that cannot send reference images to a hosted service.
Why it is compelling
- No per-model fees. You pay only for your own compute.
- Permissive code licence. The repository is under the MIT License.
- Full training codebase. Train from scratch or fine-tune on custom datasets.
- Rich materials. Roughness, metallic and opacity maps, including transparency support.
- Try before installing. Microsoft runs a public Hugging Face Spaces demo.
What it demands
- Images only. There is no text-to-3D; pair it with an image generator.
- Serious hardware. Tested only on Linux, with an NVIDIA GPU of at least 24 GB. Microsoft verified it on A100 and H100.
- An involved setup. CUDA Toolkit (12.4 recommended), Conda and several compiled packages (flash-attn, nvdiffrast, nvdiffrec, CuMesh, O-Voxel, FlexGEMM).
- Heavy meshes. The example export targets about 1,000,000 faces with 4096 textures, so decimate for games and the web.
- Manual transparency. GLB exports are opaque by default; you connect the alpha channel yourself.
- Dependency licences. Components such as nvdiffrast carry their own licences, and the weights have their own model card. Check both before you ship.
TRELLIS.2 pricing
Free to download and run on your own hardware. Your real costs are GPU time, setup and cleanup. To run open 3D models on rented GPUs, see our sister guide to the best inference providers.
Best TRELLIS.2 alternatives
- Tencent Hunyuan3D: open 2.1 weights plus a hosted pay-as-you-go API.
- Spline: runs Trellis 2 as a hosted engine inside its editor.
- Meshy: hosted, with text input, rigging and animation.
TRELLIS.2 FAQ
Is TRELLIS.2 free for commercial use?
The TRELLIS.2 code is released under the MIT License. Check the Hugging Face model card for the weights, and the licences of each dependency, before commercial use.
What GPU does TRELLIS.2 need?
An NVIDIA GPU with at least 24 GB of memory on Linux. Microsoft verified it on A100 and H100 GPUs.
Can TRELLIS.2 generate from text?
No. TRELLIS.2 is an image-to-3D model. Generate a reference image first, then convert it.
Sources
- TRELLIS.2 on GitHub
- TRELLIS.2 MIT License
- TRELLIS.2-4B weights
- TRELLIS.2 Hugging Face demo
- TRELLIS.2 paper
Independent research, updated 9 November 2026. No hands-on testing claims.
← Back to the ranked list of the best AI 3D modelers · Pricing comparison · Methodology
Named on this page as a provider? Join the network for a linked profile, Visit buttons and featured placement.