Meshy
AI 3D model generator that turns text prompts or images into textured, riggable meshes in about a minute.
Hunyuan3D
Tencent's AI 3D generator with open weights for self hosting and a commercial cloud API, strongest on hard surface geometry.
AI 3D Generation Open Source
Overview
Hunyuan3D is Tencent's generative 3D family, run on a deliberately split model: the 2.0 and 2.1 generations are published with full weights and training code for self hosting, while the newest engines (2.5, 3.0 and 3.1, plus the PolyGen topology optimizer) stay closed behind the Tencent Cloud API. Generation is two stage, a flow based diffusion transformer builds the shape and a separate Paint model synthesises PBR textures, and its standout strength versus rivals is hard surface work: flat planes, symmetry and clean seams between parts. Tencent launched the 3D engine globally in November 2025, citing more than 150 enterprise customers including Unity China and Bambu Lab. The open weights come with real strings attached though: the community license excludes the EU, UK and South Korea, caps free commercial use at one million monthly active users, and forbids using outputs to train competing models.
Key features
Two stage shape and texture pipeline
A flow based diffusion transformer generates geometry, then the separate Hunyuan3D-Paint model synthesises albedo, roughness, metallic and normal maps. Because the stages are decoupled, Paint can also texture meshes you modeled yourself.
Hard surface fidelity
Its distinctive strength in head to head tests: flatter flat surfaces, better symmetry and cleanly defined seams between sub objects than rivals, which makes mechanical and architectural generations far easier to post process.
Open weights for self hosting
The 2.1 release ships a 3.3 billion parameter shape model, texture model and training code. Teams run it locally or in ComfyUI for full data privacy and zero per generation cost, within the community license terms.
Multi view reconstruction
The commercial 3.1 API accepts up to eight views of an object (front, back, sides, top and corners), sharply reducing the guessed geometry that single image generation invents on unseen surfaces.
High resolution PBR on the API
The closed 3.0 and 3.1 engines run 10 billion plus parameter models with 4K and 8K texture output and watertight meshes aimed at 3D printing; self hosted texture resolution is bounded by your VRAM.
The HunyuanWorld ecosystem
A sibling open project extends the stack from single assets to explorable 360 degree scene generation, signalling where Tencent is taking the platform next.
Pros and cons
What users consistently praise
- Best in class hard surface generation, with flat planes, symmetry and part seams that survive close inspection
- PBR texture quality that reacts correctly to light, and the Paint stage can texture your own hand made meshes
- Genuinely self hostable with full data privacy, a rarity in this category, plus a huge open ecosystem (14K+ GitHub stars, 3M+ Hugging Face downloads)
- Flexible economics: free daily web quota, cheap per model API pricing, or zero marginal cost locally
What users consistently criticise
- Organic subjects, especially human faces and hands, routinely come out warped to the point of being unusable
- Open source outputs are dense unoptimized triangles, often over 500,000 faces; a community test of game assets found only 61 percent production ready versus 78 percent for Meshy
- The community license bans the EU, UK and South Korea, caps free commercial use at 1M MAU, and forbids training competing models on outputs
- The newest engines (2.5, 3.0, 3.1 and the PolyGen optimizer) stayed closed source, a sore point for the community that expected them to follow 2.1
Recurring themes across third-party user reviews (G2, Capterra, TrustRadius).
Videos
Pricing
| Plan | Price | What you get |
|---|---|---|
| Web studio | Free | The hosted studio grants 20 free generations per day with a Tencent login, mainly as a trial environment. |
| Self hosted (2.0/2.1) | Free (your hardware) | Weights and code are free under the Tencent Hunyuan community license. Not valid in the EU, UK or South Korea, and free commercial use ends at 1M monthly active users. |
| API, Express | USD 0.30/model | Tencent Cloud pay as you go (15 credits at USD 0.02 each). PBR materials add USD 0.20; low poly topology optimization runs USD 0.60. |
| API, Professional | USD 0.50/model | The higher quality endpoint. A typical professional generation with PBR comes to about USD 0.70. |
| Prepaid packs | From USD 0.0135/credit | One year credit packages (1,000 for USD 15 up to 100,000 for USD 1,350) cut the per credit rate versus post paid billing. |
Best for
Hard surface and product work
Vehicles, machinery, architecture and props play to its measured strengths: flat planes stay flat, symmetry holds and part seams come out clean.
E-commerce digitisation at scale
Eight view photo input plus material faithful PBR textures suit turning product photography into web ready 3D viewers in bulk.
Teams that need self hosting
The only major generator whose weights you can run on your own GPUs, keeping sensitive IP in house with no per generation cost.
Researchers and tinkerers
Full training code, a huge GitHub community and ComfyUI integration make it the default base for 3D generation experiments.
Who should look elsewhere
Character artists and animators
Aggressive smoothing disfigures faces, hands and organic anatomy, and the base model outputs static meshes with no rigging (auto rigging is a separate Tencent Cloud API).
Businesses in the EU, UK or South Korea
The community license explicitly excludes these territories, so the open weights cannot legally be used there; only the cloud API remains an option.
Low spec local hardware
The official 2.1 pipeline wants about 10 GB of VRAM for shapes and 29 GB for the full PBR pipeline. Community builds lower the floor, at the cost of heavy system RAM use.
Frequently asked questions
Can I use Hunyuan3D for free in commercial work?
Yes, if you self host the open 2.0/2.1 weights, your products stay under one million monthly active users, and you are not operating in the EU, UK or South Korea, all conditions of the community license. Outside those bounds you need the paid Tencent Cloud API or a custom enterprise license.
What is the difference between Hunyuan3D 2.1 and 3.1?
2.1 is the newest version you can download and run locally: a 3.3 billion parameter shape model, free but hardware bound. 3.1 is the closed cloud API generation running 10 billion plus parameters, with eight view input, watertight meshes for printing and 4K/8K PBR textures.
What hardware do I need to self host it?
Officially, the 2.1 pipeline takes about 10 GB of VRAM for shape generation and 29 GB for shape plus textures. The community WinPortable build with memory offloading lowers that to roughly 3 to 6 GB of VRAM, but then expect it to consume 24 GB or more of system RAM.
Why does texture generation fail on my Mac?
The texture rasterizer compiles a custom CUDA extension, and Apple Silicon has no CUDA. M series Macs can generate the untextured shape via PyTorch MPS, but PBR texturing fails, so full pipelines need an NVIDIA GPU or the cloud API.
Are the models game ready out of the box?
Rarely. Self hosted output is dense raw triangles that need manual retopology and rigging before real time use. The commercial API offers a low poly topology option (about USD 0.60 per model) that gets much closer to engine ready.
Does Hunyuan3D rig characters for animation?
Not in the base generation, output meshes are static. Tencent Cloud sells a separate auto rigging API, or you rig manually in Blender or Maya. Rivals like Meshy and Tripo include auto rigging natively.
Latest official news
Alternatives to Hunyuan3D
See all AI 3D Generation ›
Tripo
AI 3D workspace by VAST that generates textured, riggable models from text, images or sketches in seconds.
Rodin (Hyper3D)
Deemos' AI 3D generator focused on high resolution quad geometry and PBR textures from text or multi image input.