TRELLIS
A solid pick for self-hosting and full data control.
Overview
Built by Microsoft, TRELLIS is an open-weight 3D generative model (1.2B parameters) that converts a single image into a structured 3D asset, and it's light enough to run on laptop-class hardware. The MIT licence covers the code, but commercial-use terms aren't fully specified, so teams should check before shipping a product on it. It has no text context window, so it's purely a 3D-asset generator, not a general assistant, and quality still trails specialised proprietary 3D pipelines on complex geometry. Self-hosted, it keeps every input and output on the user's own infrastructure regardless of the vendor's home country.
Skill profile
Not disclosed
Strengths
- Open-source and self-hostable
Limitations
- API pricing not disclosed
Who is it for
- you want to control cost or self-host
- you have GDPR constraints
Ideal use cases
- rapid 3D asset prototyping
- game asset generation
- e-commerce product visualization
- on-premise 3D pipeline
- 3D generation research
Access & availability
Key specifications
Privacy
Run it locally
| Quantization | Disk | RAM / VRAM | Typical hardware |
|---|---|---|---|
| Q4 · recommended | 0.7 GB | 3 GB | Any recent PC/Mac |
| Q8 · balanced | 1.3 GB | 3 GB | Any recent PC/Mac |
| FP16 · max quality | 2.4 GB | 5 GB | Any recent PC/Mac |
Estimates for a moderate context. Long contexts need more RAM (KV cache).
Deploy
Copy-ready commands generated from this card. Adjust context length and GPU count to your hardware.
OpenAI-compatible server for production on NVIDIA GPUs.
pip install vllm
vllm serve JeffreyXiang/TRELLIS-image-large \
--max-model-len 32768 \
--tensor-parallel-size 1 \
--dtype autoAdvanced data · for expertsArchitecture, modalities, detailed cost, full benchmarks▾
| Arena Elo | — |
| MMLU | — |
| GPQA | — |
| HumanEval | — |
| SWE-Bench | — |
| MATH | — |