Vi3W
Skip to content

Guides

Best open-source 3D AI for 2027: the models you can run, sorted by licence and GPU

Open-source 3D AI to run going into 2027: TRELLIS.2, TRELLIS, Hunyuan3D 2.x, TripoSG, TripoSR and SAM 3D, sorted by licence, VRAM and upkeep.

5 min read

An open glass book floating in the air with small 3D models growing out of its pages like a pop-up book
Illustration. Open weights are only half the story. The licence and the GPU decide what you can do with them.

Quick answer

Which open-source 3D AI should you use going into 2027?

As of November 2026, TRELLIS.2 is the newest fully permissive pick: MIT-licensed, but it wants a 24 GB NVIDIA card. On 16 GB, run TRELLIS v1, also MIT. On smaller GPUs, TripoSG or TripoSR. Hunyuan3D 2.x runs on less VRAM still, but its licence isn't valid in the EU, UK or South Korea.

At a glance

Open 3D models, from each project's own README and licence file, checked November 21, 2026. VRAM is what the README states.
ModelLicenceVRAM in the READMEWhat it doesLast activity
TRELLIS.2 (Microsoft)MIT24 GB+ NVIDIA, Linux testedImage to PBR GLB; no text modeActive 2026
TRELLIS (v1)MIT16 GB+ NVIDIA, Linux testedImage to 3DCommit Oct 27, 2026
Hunyuan3D 2.1 (Tencent)Tencent Hunyuan community licence (not valid in EU, UK, South Korea)10 GB shape, 21 GB texture, 29 GB bothImage to textured meshCommit Oct 17, 2025
Hunyuan3D 2.0 / 2miniSame licence family6 GB shape, 16 GB with textureImage to mesh; macOS, Windows, LinuxReleased Jan to Mar 2025
TripoSG (VAST)MIT8 GB+Image to mesh (1.5B)Released Mar 2025
TripoSRMITAbout 6 GBFast single image to meshReleased Feb 2024
SAM 3D Objects (Meta)SAM License (custom, gated weights)32 GB+ NVIDIA, LinuxImage plus mask to splat (mesh optional)Active to Jun 2026

Why this list sorts by licence and GPU

Most "best open-source 3D AI" lists rank by output quality. That's the wrong first question. If the licence doesn't cover your country or your client work, or the model won't fit on your card, quality doesn't matter. So this list goes in that order: licence first, VRAM second, upkeep third, then what the model makes.

These are the picks to use going into 2027, as of November 21, 2026. No one knows what opens up next year, so there are no predictions here, and we'll update the page as 2027 releases land. For the wider field, including hosted tools, see best 3D AI for 2027. For a longer evergreen tour, see open-source 3D AI.

A plain statement before the picks: Vi3W isn't open source. It hosts TRELLIS, TRELLIS 2 and Hunyuan 3D models in the browser for people who don't have the GPU, which we come back to at the end.

The licence tiers

Permissive: MIT and Apache-2.0

You can use these commercially, modify them and ship them, as long as you keep the licence notice. TRELLIS, TRELLIS.2, TripoSR, TripoSG and UniRig are MIT. InstantMesh, Step1X-3D and DeepMesh are Apache-2.0.

Custom but royalty-free

Meta's SAM License (used by SAM 3D Objects and Body) is royalty-free and has no non-commercial clause, but it's a custom licence with conditions: pass it on when you redistribute, acknowledge SAM in publications, no military or ITAR uses. Check with a lawyer before you call it "commercial use OK". Stability AI's Community Licence (Stable Fast 3D, SPAR3D) is free under US$1M annual revenue, with registration for commercial use.

Region-restricted

The Tencent Hunyuan community licence covers Hunyuan3D 2.0, 2mini and 2.1. It isn't valid in the EU, the UK or South Korea. If you work there, these models are off the table however good they are.

Non-commercial research

Plenty of interesting 3D research is non-commercial: MeshAnything V2 (S-Lab License), Text2CAD (CC BY-NC-SA 4.0), CAD-Recode (CC BY-NC 4.0). Fine for learning, not for client work. Details: open-source 3D AI licences.

A rooftop garden of glowing 3D models growing in planters, tended by small robot gardeners, with an open gate in the wall
Illustration. An open garden still has rules about what you can take home.

The picks, by GPU

24 GB or more: TRELLIS.2

TRELLIS.2 is Microsoft's 4B-parameter model, MIT-licensed. It turns one image into a PBR GLB with base colour, roughness, metallic and opacity maps. It has no text mode. Officially it needs an NVIDIA GPU with at least 24 GB and has only been tested on Linux.

Two things changed in 2026 that make it easier to run. ComfyUI made TRELLIS.2 a core node in v0.34 (August 2026), with no compiled CUDA extensions, which helps Windows users. And community GGUF builds exist for people chasing smaller cards; see TRELLIS 2 on 16 GB of VRAM.

On the 3D Arena community leaderboard, TRELLIS.2-4B ranked 18th on November 21, 2026, below TRELLIS v1 (6th) and TRELLIS-3DGS (2nd). Votes on newer entries are thin, so test it on your own images.

16 GB: TRELLIS v1

The original TRELLIS is MIT, needs 16 GB of NVIDIA VRAM per its README and is still maintained, with a commit on October 27, 2026. It's the safest permissive choice on a 16 GB card. Setup: run TRELLIS locally.

8 GB or so: TripoSG and TripoSR

TripoSG (VAST, 1.5B, March 2025) is MIT and asks for 8 GB or more. TripoSR (February 2024) is MIT and runs in about 6 GB. Both are older than the TRELLIS models, but they're the permissive options for mid-range cards.

Low VRAM, if the licence fits: Hunyuan3D 2.x

Hunyuan3D 2.0 needs about 6 GB for shape and 16 GB with texture, and runs on macOS, Windows and Linux. Hunyuan3D 2.1 (June 2025) needs 10 GB for shape, 21 GB for texture or 29 GB for both, with a --low_vram_mode. It's still Tencent's newest open asset model. The newer HY 3D 3.0 and 3.1 are API and web only, not open weights.

Workstations: SAM 3D Objects

Meta's model takes an image plus a mask, so it can lift one object from a busy photo. It needs Linux and a 32 GB NVIDIA GPU, outputs a Gaussian splat by default (a mesh needs a code flag), and the weights are gated on Hugging Face.

Beyond generation

  • Rigging: UniRig (MIT, SIGGRAPH 2025) rigs humans, animals and objects. Its successor SkinTokens (MIT) does skeleton and skinning in one model.
  • Artist-style meshes: DeepMesh is Apache-2.0. MeshAnything V2 is non-commercial.
  • Worlds: Tencent's HY-World 2.0 outputs meshes and 3DGS worlds, open-sourced in parts, under a custom licence (check).
  • Unmaintained: OpenAI's Shap-E hasn't been updated since November 2023. Stability AI hasn't released a new 3D model in 2026.
  • The open line is drawn at older versions. Tencent's newest asset models (HY 3D 3.x) are closed, while 2.1 stays open.
  • Open world models are arriving, with real exports. HY-World 2.0 outputs meshes and splats, unlike explore-only world models.
  • Rigging and animation go open too. Tencent's HY-Motion 1.0 is open-source text-to-motion, with FlowHMR video-to-motion (.fbx) built on it in October 2026.
  • Churn. Check the repo's last commit before you build a pipeline on it.

When you don't have the GPU

If none of these fit your card, Vi3W runs TRELLIS (the free default), TRELLIS 2 and Hunyuan 3D in the browser, along with Meshy, Tripo and Rodin, then lets you rig, retexture or retopologise the result. It's hosted, not open source. Try Vi3W.

Frequently asked questions

best ai 3d model generator open source

TRELLIS.2 if you have a 24 GB NVIDIA GPU, since it's MIT-licensed and writes PBR GLBs. TRELLIS v1 on 16 GB. Hunyuan3D 2.x if your GPU is smaller and you're outside the EU, UK and South Korea, where its licence doesn't apply.

ai 3d model generator free open source

The weights cost nothing, but you pay in hardware and setup time. TRELLIS, TRELLIS.2, TripoSR and TripoSG are MIT. Run them in ComfyUI, which has supported TRELLIS.2 natively since v0.34, or try the TRELLIS.2 Hugging Face Space first.

ai 3d generator open source

Open 3D generators are image-first. Of the current ones, TRELLIS.2 has no text mode at all, so most open text-to-3D setups make an image first and then turn it into a mesh.

Last reviewed