Guides
Meshy alternatives you run locally, matched to the GPU you have
Meshy alternatives you can run locally: TRELLIS.2, TRELLIS, Hunyuan3D 2.x, TripoSR and more, matched to your GPU's VRAM, licence and OS.
5 min read

Quick answer
Which Meshy alternative can you run on your own computer?
With a 24 GB NVIDIA card, TRELLIS.2: MIT-licensed, PBR GLB output and native ComfyUI nodes. With 16 GB, TRELLIS or Hunyuan3D 2.1's low-VRAM mode. With 6 to 8 GB, Hunyuan3D 2mini for shape, TripoSR or TripoSG. Meshy itself has no local version; its plugins connect to Meshy's servers.
At a glance
| Model | Licence | VRAM per README | Notes |
|---|---|---|---|
| TRELLIS.2 (Microsoft) | MIT | NVIDIA, 24 GB or more | Linux tested; native ComfyUI nodes since v0.34; PBR GLB out; community builds aim lower (check) |
| TRELLIS (Microsoft) | MIT | NVIDIA, 16 GB or more | Linux tested; Windows not fully tested; still maintained |
| Hunyuan3D 2.0 / 2mini | Tencent Hunyuan community licence | 6 GB shape; 16 GB shape and texture | macOS, Windows, Linux; licence not valid in the EU, UK or South Korea |
| Hunyuan3D 2.1 | Tencent Hunyuan community licence | 10 GB shape; 21 GB texture; 29 GB both | Has a low-VRAM mode; same regional limits |
| TripoSR | MIT | About 6 GB | From Stability AI and Tripo, February 2024 |
| TripoSG (VAST) | MIT | 8 GB or more | 1.5B parameters, March 2025 |
| Stable Fast 3D | Stability AI Community Licence | About 6 GB | Free under US$1M annual revenue; Mac via MPS is experimental |
| SAM 3D Objects (Meta) | SAM License, gated weights | 32 GB or more | Linux; outputs a Gaussian splat by default |
Why go local, and what it really costs
Meshy runs on Meshy's servers. Its web app, its API and its Blender and ZBrush plugins all connect to that service, so there's no version you can install and run offline. If you want generation without credits, without uploading your images and without an account, you need an open model on your own hardware.
That trade is real in both directions. You stop paying per model and your reference images stay on your machine. In return, you take on the setup, the VRAM limits and the parts of Meshy that aren't the model: rigging, animation, retexturing, printing fixes and a choice of export formats. This page is organised around the GPU you actually have, because that decides your options more than anything else.
If you mainly want the list of open models and their licences, start with open-source Meshy alternatives. This page is about running them.
If you have a 24 GB NVIDIA card
Use TRELLIS.2. Microsoft's 4-billion-parameter model is MIT-licensed and turns a single image into a PBR GLB with base colour, roughness, metallic and opacity. Its README lists an NVIDIA GPU with at least 24 GB, verified on A100 and H100 cards, and testing on Linux only. It does image to 3D only; there's no text input.
The big change in 2026 is ComfyUI. Since v0.34 (August 2026), TRELLIS.2 is a core ComfyUI node. It needs no custom nodes and no compiled CUDA extensions, because the post-processing was rewritten in PyTorch and SciPy. That also removed two dependencies that carried non-commercial licences. Comfy-Org publishes official repackaged weights, including an int8 diffusion model. On Windows, that's now by far the easiest route.
One export detail: TRELLIS.2 writes its GLB in opaque mode by default. The alpha stays in the texture, but you need to wire it to the material's alpha in your 3D app. Our TRELLIS 2 in ComfyUI guide walks through the nodes.
If you have 16 GB
TRELLIS (the original) lists 16 GB or more and is still maintained, with commits as recently as October 2026.
Hunyuan3D 2.1 needs about 10 GB for shape, 21 GB for texture and 29 GB for both, and has a --low_vram_mode flag. At 16 GB, expect to generate shape locally and texture some other way.
TRELLIS.2 on 16 GB isn't official. Secondary guides say the heavier step needs about 16 GB at lower resolution and suggest starting at 512, with 24 GB recommended for 1024 and above. Community GGUF builds exist, such as ilintar/trellis2-gguf for the trellis.cpp runtime and Aero-Ex/Trellis2-GGUF for ComfyUI's GGUF nodes. Treat them as experiments: results vary.

If you have 6 to 8 GB
You can still make shapes. Hunyuan3D 2.0 and 2mini list about 6 GB for shape alone and 16 GB with texture, and run on macOS, Windows and Linux. TripoSR, from Stability AI and Tripo, is MIT-licensed and needs about 6 GB. TripoSG from VAST is MIT and needs 8 GB or more. Stable Fast 3D needs about 6 GB, but its Stability AI Community Licence is only free under US$1 million in annual revenue, and you must register for commercial use.
These are older or smaller models, so compare them against a hosted model on the same image before you rely on them.
If you're on a Mac
There's no official Apple Silicon support for TRELLIS.2. Community ports exist: gtrg55/trellis2-mlx (MIT, native MLX, tested on an M4 Max with 128 GB), lyonsno/trellis2mlx, sanchez-kim/trellis-silicon and a ComfyUI-TrellisMac node. All are small hobby projects, so treat them as experimental. Hunyuan3D 2.0 lists macOS support, and Stable Fast 3D runs on Mac through MPS experimentally. For more, see Hunyuan 3D on Mac.
Licences: free to run isn't always free to use
- MIT (TRELLIS, TRELLIS.2, TripoSR, TripoSG): use it commercially, keep the licence notice.
- Tencent Hunyuan community licence (Hunyuan3D 2.x): not valid in the EU, UK or South Korea. Tencent's newer HY 3D 3.x models are web and API only, not open weights.
- Stability AI Community Licence (SF3D): free under US$1M annual revenue.
- SAM License (SAM 3D Objects): royalty-free with no non-commercial clause, but gated weights and other conditions. It also needs a 32 GB card and outputs a Gaussian splat by default, so it's a poor fit for a typical Meshy replacement.
More detail in open-source 3D AI licences.
Filling the gaps Meshy covers
A local model gives you a mesh. Meshy also rigs, animates, retextures and preps for printing. Local stand-ins:
- Rigging: UniRig (MIT) handles humans, animals and objects; its successor SkinTokens (MIT) does skeleton and skinning in one model. Blender's bundled Rigify is free but not AI.
- Retopology: Blender's QuadriFlow remesh, or Instant Meshes, which is free and unmaintained since 2022 but still works.
- Texturing: StableGen, a free GPL-3.0 Blender add-on that textures meshes through ComfyUI, and Hunyuan3D-Paint 2.1, which needs 21 GB.
Each is its own install, so budget setup time.
When local isn't worth it
If your GPU is below 6 GB, you're on a Mac without much memory, or you just want a model by this afternoon, a hosted tool is the honest answer. That's where Vi3W fits, and to be clear: Vi3W is not a local tool. It runs in the browser, on our servers, with no GPU needed. It includes TRELLIS as the free default and 24 other image-to-3D models, including Meshy, Tripo, Rodin and Hunyuan 3D. It also covers the gaps above with Add bones, Make it move, New paint job and Clean shape, and downloads in GLB, FBX, OBJ, STL, PLY and glTF. Try Vi3W.
Frequently asked questions
meshy alternative local
TRELLIS.2 is the strongest local option if you have an NVIDIA GPU with 24 GB. It's MIT-licensed, outputs PBR GLB and runs as native ComfyUI nodes. On smaller cards, Hunyuan3D 2.0 and 2mini need about 6 GB for shape, and TripoSR about 6 GB.
meshy ai alternative free download
The free downloads are open model weights, not apps. TRELLIS and TRELLIS.2 (MIT) come from Microsoft's GitHub and Hugging Face, with official ComfyUI weights at Comfy-Org/TRELLIS.2. Hunyuan3D 2.x is free too, but its licence excludes the EU, UK and South Korea.
Last reviewed

