Vi3W
Skip to content

Guides

Open-source text to 3D model AI: which open projects take a prompt in late 2026

Which open-source AI turns text into a 3D model in late 2026: TRELLIS v1, the text-to-image-to-3D route, code-CAD with build123d, and what to skip.

5 min read

A framed photograph of a teapot, with the same teapot stepping out of the frame as a solid 3D object
Illustration. In open source, most text-to-3D still passes through a picture on the way.

Quick answer

Is there open-source AI that makes a 3D model from text?

Yes, but fewer projects than you'd expect take a prompt directly. As of November 2026, TRELLIS v1 (MIT) lists text-conditioned models, while TRELLIS.2 is image-only and Shap-E is unmaintained. The dependable open route is text to image to 3D, or, for mechanical parts, an LLM writing build123d or CadQuery code.

At a glance

Open projects that turn text into 3D, from each project's own README and licence file, checked November 21, 2026.
ProjectLicenceTakes text directlyOutputStatus
TRELLIS (v1)MITText-conditioned models listed (check the README)3D meshMaintained; commit Oct 27, 2026; 16 GB NVIDIA
TRELLIS.2MITNo, image onlyPBR GLBActive 2026; 24 GB NVIDIA
Shap-E (OpenAI)MITYes3D shapesUnmaintained since Nov 2023
HY-Motion 1.0 (Tencent)Open weights (check licence)Yes, text to human motionMotion; FlowHMR adds .fbx from videoWeights Dec 30, 2025
build123dApache-2.0Via code an LLM writesSTEP and other CADActive Nov 2026
Text2CADCC BY-NC-SA 4.0 (non-commercial)YesCAD sequencesLast push May 2025

The short version

Search for open-source text to 3D and you'll find plenty of demos. Check what each project actually accepts and you'll find a surprise: the best open 3D models of 2026 are image models. Microsoft's TRELLIS.2, the newest permissive release, has no text mode at all.

That leaves three honest routes from a prompt to an open-source 3D model:

  1. A model that takes text directly. Few are current.
  2. Text to image, then image to 3D. Two open models chained together.
  3. Text to code, then code to geometry. An LLM writes CAD code that an open library runs.

This page goes through each, as of November 21, 2026. For the full how-to on route two, see the companion guide open-source text to 3D.

Route 1: open models that take text directly

TRELLIS (v1)

The original TRELLIS is MIT-licensed and still maintained, with a commit on October 27, 2026. Its repository lists text-conditioned models alongside the image ones; check the README for which checkpoints are current before you download. It needs an NVIDIA GPU with 16 GB or more and is tested on Linux. Setup: run TRELLIS locally.

Shap-E

OpenAI's Shap-E (MIT) takes text or images. Its last commit was November 8, 2023, and it hasn't been updated since. It isn't archived, but don't build anything on it in 2026.

What about Hunyuan3D and TRELLIS.2?

TRELLIS.2 is image to 3D only. The Hunyuan3D 2.x repositories are built around image input, so plan on an image step first and check the README for any text pipeline in your version. Their licence also isn't valid in the EU, the UK or South Korea. The newer HY 3D 3.x models, which do take text, are API and web only, not open weights.

If what you want from text is movement rather than shape, Tencent's HY-Motion 1.0 is open-source text-to-3D human motion, with weights published December 30, 2025. FlowHMR, built on it in October 2026, turns video into .fbx motion on a 12 GB+ consumer GPU. Check its licence before commercial use.

An open glass cube revealing glowing gears inside, with a small sculpted bird being put together among them
Illustration. Open source lets you see the gears. It doesn't always give you a text box.

Route 2: text to image to 3D

This is the route most open setups use, and for good reason. You see the design as an image before you spend GPU time on 3D, and you can throw away bad images for almost nothing.

  1. Prompt an open image model for one object, centred, on a plain background, three-quarter view, even light.
  2. Make several and pick the cleanest.
  3. Feed it to an image-to-3D model. TRELLIS.2 (MIT, 24 GB, PBR GLB), TRELLIS v1 (MIT, 16 GB), TripoSG (MIT, 8 GB+) or TripoSR (MIT, about 6 GB).
  4. Clean up in Blender: scale, decimate or retopologise.

ComfyUI is where people chain this. Since v0.34 (August 2026), TRELLIS.2 is a core node with no compiled CUDA extensions, plus mesh post-processing, UV unwrap and normal and AO baking nodes. See ComfyUI 3D.

The licence trap: two models means two licences. A permissive 3D model doesn't help if the image model forbids commercial use. Check both in open-source 3D AI licences.

Route 3: text to CAD code

For brackets, enclosures and anything with dimensions, a mesh generator is the wrong tool. Ask an LLM to write code for an open CAD library instead:

  • build123d (Apache-2.0, active November 2026) and CadQuery (Python, built on OCCT, active November 2026) both export STEP.
  • OpenSCAD and Blender Python are the other common targets. ChatGPT and Claude can write both, but neither makes meshes natively.

There's also open research that turns text into CAD directly: Text2CAD (NeurIPS 2024 Spotlight), but it's CC BY-NC-SA 4.0, so it's non-commercial. More in open-source text to CAD.

Which route for which job

  • Quick experiments with one prompt: TRELLIS v1's text models.
  • Props, characters and anything with a look: text to image to 3D, so you control the design.
  • Parts with real measurements: text to code with build123d or CadQuery.
  • Animation from a description: HY-Motion 1.0.

Hardware for each route

The routes ask very different things of your computer:

  • Route 1 needs the 3D model's VRAM: 16 GB of NVIDIA memory for TRELLIS v1, per its README.
  • Route 2 needs an image model and a 3D model. You can run them one after the other so they don't share memory, but the 3D step still sets the floor: 24 GB for TRELLIS.2 officially, less for TripoSG or TripoSR. Community GGUF builds of TRELLIS.2 aim lower, with mixed results; see TRELLIS 2 on 16 GB of VRAM.
  • Route 3 runs on almost anything. The LLM can be hosted, and CAD libraries like build123d run on an ordinary laptop.

On a Mac, routes 1 and 2 are hard, because these models target NVIDIA GPUs. Route 3 works fine.

How to test an open text-to-3D setup

Pick three prompts: a hard-surface prop, a creature and a thin object such as a sword. Run each through your route, then check the wireframe in Blender, the texture maps, the face count and the licence of every model in the chain. Don't trust the first result: run each prompt more than once.

When you'd rather not run it yourself

Vi3W isn't open source, but it takes text directly and also runs TRELLIS, TRELLIS 2 and Hunyuan 3D models in the browser, with no GPU. Then you can rig, animate, retexture or retopologise the result and download it as GLB, FBX, OBJ, STL, PLY or glTF. Try Vi3W.

Frequently asked questions

best text to 3d model ai open source

For direct prompts, TRELLIS v1: it's MIT, maintained in 2026 and lists text-conditioned models. For better control, generate an image first and feed it to TRELLIS.2, which is image-only. Check both licences in the chain.

best free text to 3d model

Free in money terms: an MIT model on your own GPU. Free without a GPU: hosted free tiers, such as Meshy's 100 credits a month (CC BY 4.0 output). Each has a cost somewhere, either hardware and setup or limits and attribution.

text to 3d model ai free download

The open weights for TRELLIS and TRELLIS.2 are downloadable from the projects' GitHub and Hugging Face pages. ComfyUI ships TRELLIS.2 nodes natively from v0.34, with official repackaged weights at Comfy-Org/TRELLIS.2.

Last reviewed