TRELLIS.2 Download: Official Weights and Safe Setup

Download from the official source

The TRELLIS.2 model is published by Microsoft Research. Use the official repository and model card rather than a re-uploaded archive:

• Microsoft TRELLIS.2 repository
• TRELLIS.2-4B model on Hugging Face

trellis-2.com is an independent hosted service. It does not redistribute the model weights, and a link to an upstream project is not an endorsement of every third-party wrapper or workflow that uses it.

Read the licence before the download

The official repository states an MIT licence for TRELLIS.2. That covers the code and weights under the terms in the repository, not your source images, your generated output, or a community wrapper's separate licence. Check all three before commercial use.

Minimum setup to expect

The official README currently describes:

• A 4B image-to-3D model using O-Voxel
• Linux and an NVIDIA GPU with at least 24 GB of memory
• Verification on A100 and H100 GPUs
• Python 3.8 or later, Conda, PyTorch 2.6.0, and CUDA 12.4
• flash-attn by default, with xformers as the documented fallback

These requirements can change. Read the README at the exact commit you plan to install rather than copying an old command from a tutorial.

Safe download and install order

1. Open the official repository and note the commit you intend to use.
2. Clone recursively and follow the current setup instructions.
3. Download the model from the linked model repository, not a mirror.
4. Record the model revision and file hashes.
5. Run the official example before adding ComfyUI or another wrapper.
6. Only then review a community wrapper's code, licences, and dependencies.

The full local setup checklist covers CUDA, memory, and validation. The ComfyUI guide separates the official project from community integrations.

What you are not downloading

You are not downloading a hosted account, a managed GPU, or a finished asset pipeline. The weights do not remove the need for:

• A compatible GPU and driver stack
• Storage for models and temporary render files
• A security review of custom nodes and binary wheels
• Scale, topology, UV, and material validation in a 3D tool
• A licence review for source images and generated output

If you want to evaluate one image without installing the stack, use the hosted image-to-3D generator.

FAQ

Is the download free?

The official model is published as open source under the licence in Microsoft's repository. Your hardware, electricity, and setup time are still real costs.

Can I use the official weights on Windows?

The official README currently documents Linux and verified A100/H100 support. Do not assume that a consumer Windows machine will work without checking the current compatibility notes and testing your own environment.

Why not download from a mirror?

A mirror can be stale, modified, or unlicensed. Use the official repository and model card, then record the revision and hashes you installed.

Can I try the model without downloading it?

Yes. The hosted generator is the shorter path for one evaluation. It shows the credit cost before submission, and it returns a GLB you can inspect in the browser viewer.

Open the official Microsoft repository

Read the local setup checklist