Meshy
Beijing engineering team, Santa Clara registration, founded by Taichi author Yuanming Hu. Meshy 6 reached general availability on 18 January 2026 with meshes to roughly 600K faces.
Overview
Origin first, because it determines how you file this one. Yuanming Hu graduated from Tsinghua's Yao Class in 2017, took a PhD at MIT and wrote Taichi, a language for high-performance graphics computation that a large amount of research code now runs on. He founded Taichi Graphics in Beijing in 2021 and pivoted the team to Meshy in 2023. The company today runs about 150 people split between a Santa Clara registration and a substantial Beijing engineering operation, and Chinese trade press treats it as a Beijing company while the corporate paperwork does not. Both descriptions are accurate, which matters if your procurement policy keys on one of them.
Meshy 6 reached general availability on 18 January 2026 after an October 2025 preview, aiming at sculpting-level fidelity with cleaner topology on hard-surface parts as well as organic shapes, and meshes up to roughly 600K faces. The release added a Low Poly Mode for real-time budgets, multi-colour 3D printing output and new text and image endpoints. Credit costs are unusually legible for this category: 20 credits for an untextured Meshy-6 generation, 30 textured, 10 for a texture task and 5 for a remesh, which puts a finished game-ready character near 35 credits. Plans run free, $20 Pro, $40 Premium, $100 Ultra and $70 per seat for Studio, with roughly 20 percent off on annual billing. One useful detail for API users: should_remesh now defaults to false on Meshy-6, so retopology runs when you ask for it rather than silently on every job.
The engineering trade-off is closure. There are no weights, no local inference and no on-prem path at any price below an enterprise conversation, so anything air-gapped is out immediately. What you get in exchange is stability and clean commercial terms with no user cap and no territorial exclusion. On geometry, apply the same discipline as with every generator in this category: 600K faces of generated surface is a concept asset and a visualisation asset, not engineering geometry. Chamfers, fillets and bolt holes come back approximated, edge flow follows the surface rather than the design, and there is no parametric history to edit. Revenue is around $30 million ARR going into GDC 2026, roughly doubled in three months, with a reported Series B of close to $400 million at a stated $1.5 billion valuation behind it.
Key Features
- ✓ Meshy 6, generally available since 18 January 2026 after an October 2025 preview, targeting cleaner topology on hard-surface as well as organic assets at up to roughly 600K faces
- ✓ Text-to-3D, image-to-3D and multi-image-to-3D endpoints with PBR texturing, retopology and automatic rigging in one pipeline
- ✓ Published per-operation credit pricing: 20 credits untextured, 30 textured, 10 per texture task, 5 per remesh
- ✓ Low Poly Mode for real-time budgets and multi-colour 3D printing output, both added with the 6 release
- ✓ should_remesh now defaults to false on Meshy-6, giving you explicit control over when topology cleanup runs
- ✓ Plans from free through $20 Pro, $40 Premium, $100 Ultra and $70 per seat for Studio teams
Where it holds
- • The most predictable API in generative 3D, which is a compliment. Stable endpoints, published per-operation credit costs and SDK behaviour that does not shift underneath you between releases
- • Commercial terms are simple on every paid tier, with no monthly-active-user cap and no territorial carve-out, which several open-weight Chinese 3D models cannot claim
- • Revenue says it is a business rather than a demo: about $30 million ARR going into GDC 2026, roughly doubled in three months
- • Yuanming Hu's Taichi background shows in the systems work. Throughput is high and the failure modes are consistent rather than random
Where it breaks
- • Nothing is downloadable. No weights, no local inference and no on-prem option, which rules it out of air-gapped, classified or export-sensitive work
- • The Chinese-origin label is genuinely ambiguous. The company is registered in Santa Clara with roughly 150 staff split between the US and a substantial Beijing operation, so procurement rules written around country of incorporation will classify it differently from the engineering reality
- • Topology in 6 is cleaner but still generated rather than authored. Hard-surface parts return with wavy edge flow around chamfers and holes that a modeller has to correct before anything is manufacturable
- • Credits are spent per attempt, not per usable result, so iterating on an awkward prompt burns budget fast
- • No B-rep, no parametric output and no dimensional control, so it belongs in concepting and visualisation rather than engineering release
My Take
Cleanest API in generative 3D, and the least interesting licence. Meshy 6 has been generally available since 18 January 2026, tops out near 600K faces and charges 30 credits for a textured generation, which makes cost modelling trivial in a category that usually fights you on that. Worth knowing before you file it: this is a Beijing engineering team with a Santa Clara registration, so the Chinese-origin label describes the people rather than the paperwork. Nothing is downloadable, and that is the binding constraint if you need offline or on-prem inference.
Quick Info
- Pricing:
- freemium
- Openness:
- Proprietary
- Starting at:
- Free tier at $0, Pro at $20 a month, Premium at $40, Ultra at $100, Studio for teams at $70 per seat with a shared credit pool and enterprise on custom terms. Annual billing saves roughly 20 percent against monthly. API credit costs are published per operation: 20 credits for an untextured Meshy-6 generation, 30 for a textured one, 10 per texture task and 5 per remesh, so a finished game-ready character lands near 35 credits.
- Added:
- Aug 2026
- Updated:
- Aug 2026
Use Cases
Judge it on your own work
The notes above say where Meshy holds and where it breaks. The fastest check is your own workload.
Visit website ↗Alternatives to Meshy
Tripo
freemiumImage and text to 3D from VAST in Beijing. Hosted Tripo 3.0 Ultra reaches 2 million polygons, while TripoSG 1.5B, TripoSR and UniRig ship as MIT-licensed weights you can run yourself.
Hunyuan3D 2.1
freeTencent's open image-to-3D model with PBR texture synthesis and full training code. A separate artifact from the Hunyuan chat product, and not under an OSI-approved licence.
Rodin (Hyper3D)
freemiumDeemos (影眸科技), a ShanghaiTech spinout, builds Rodin. Gen-2.5 emits native quad meshes at 4k to 50k quads, which is why it survives real pipelines. Hosted only, no weights.