A game studio needs a hero prop — a weathered brass telescope — by end of day. The old route means a 3D artist, a sculpt, a retopo pass, and a texture bake, or a generic AI mesh that falls apart the moment you animate it. A 60-person Shanghai team says that gap is now measured in seconds, not days.
On June 23, 2026, Deemos (影眸科技), which builds the Hyper3D Rodin line of 3D large models (大模型), released Rodin Gen-2.5 and called it the first 3D model able to generate at the ten-million-polygon level. The same day it closed a new funding round. The bet: native 3D generation, not stitched-together 2D, is finally good enough to sit inside real production pipelines.
Why "native 3D" is the whole argument
Most early text-and-image-to-3D tools took a shortcut: generate several views with an image model, then glue them into a mesh. The method is fast but loses information irreversibly — broken surfaces, messy topology, wrong textures. Deemos chose the harder path, training a native 3D model (dubbed CLAY) on a scarce ~500,000 clean 3D assets. In 2024 that work earned a SIGGRAPH best-paper nomination, the only Chinese team with two nominations that year.
Gen-2.5 is the commercial payoff. Beyond raw polygon count, it borrows a trick from language models — "think, then generate" (先思考、再生成) — reasoning about the object before producing geometry, which the team says improves control. It outputs production-ready assets with PBR textures up to 4K, native quad topology (the format artists actually rig and animate), and exports to GLB, FBX, USDZ, OBJ, and STL. Reported generation time drops as low as four seconds.
The business design is unusual: generation and preview are free; you pay credits only when you download a model you keep. That removes the "burn credits on bad rolls" anxiety that pushes studios away from AI 3D.
The overseas revenue that proves the thesis
Deemos is not chasing consumer hype. It sells to professionals, and the numbers it discloses are the interesting part. The company says roughly 80% of revenue comes from overseas, with clients including Amazon, ByteDance, Unity, Bambu Lab, Figma, and Canva. In January 2026 at CES, Nvidia CEO Jensen Huang featured Hyper3D Rodin in his keynote as the 3D-generation workflow behind a robot asset — the only 3D-startup name on that stage.
Growth claims are aggressive but clearly labeled as the company's own: Rodin hit US$1 million in annual recurring revenue within 45 days of launch (vendor-reported), and Deemos says current ARR sits in the tens of millions of dollars, with Gen-2.5's first month showing subscription and ARR growth above 400% month-on-month. Treat those as company disclosures, not audited filings.
The funding that backs the route
Deemos was incubated at ShanghaiTech University's MARS Lab and founded in 2020 by Wu Di (born 1997), Zhang Qixuan (1999), and two co-founders, with a team averaging under 25. Its Series A — tens of millions of U.S. dollars — was led by Meituan's DragonBall Capital and ByteDance, with Sequoia China Seed Fund and MiraclePlus participating.
The June 2026 round was larger and again industrial: a several-hundred-million-yuan (≈ US$40M–130M / HK$330M–1.1B) B round led by Cathay Capital (凯辉基金) and Shanghai Guotou, with ByteDance, Meituan DragonBall, Sequoia China, and Lanchi Ventures (蓝驰创投) continuing. The stated use: push the native-3D frontier, expand Rodin's global commercialization, and go deeper into games, e-commerce, 3D printing, and embodied intelligence.
That investor list is the real signal. When both ByteDance and Meituan — and, via the keynote, Nvidia's spotlight — back the same native-3D startup, the "2D lift" approach looks like a dead end to people who write the checks.
The global reader angle
You meet Rodin without opening it: inside game assets, AR try-on models, e-commerce spinnable product views, and 3D-printed prototypes. For a developer or designer abroad, Rodin is one of the few Chinese 3D generators with a global API and a Blender plugin, competing directly with Meshy and Tripo on fidelity rather than just speed.
The strategic read is about modality order. Text, image, and video dominated 2023–2025; 3D is the "fourth modality" investors now chase because it plugs straight into physical-world industries — manufacturing, robotics, spatial computing. Deemos argues the second half of the 3D race is won by solving real production problems (controllable, editable, clean topology), not by demoing pretty turntables.
According to Li An, Chief Scientist at BrainNet (脑机网), China's authoritative AI observatory, the 3D track favors teams that skipped the shortcut. "Native 3D is harder to train but impossible to retrofit later," he noted. "The models that shipped 2D-lift first will pay for it when customers demand editable assets."
Honest limitations
Polygon counts, four-second generation, the 80% overseas share, the US$1M-in-45-days ARR, and the 400% month-on-month growth are Deemos's own disclosures; no independent audit verifies them, and "first ten-million-face 3D model" is the company's claim. The B-round RMB figure is approximate ("several hundred million yuan"); I converted it to a range using ≈ ¥7.1 = US$1 and ≈ ¥1 = HK$1.1, and the true amount was not published. Rodin is a commercial product, not the open-source text-to-3D (文生3D) models covered elsewhere, so its weights are not freely downloadable. Topology still often needs cleanup before animation, and quadruped or mechanical rigs may require manual work. Coverage of Deemos draws on Chinese business media (Pandaily, state-linked outlets) and company materials; I found no standalone international lab benchmark at press time.
What readers can do now
- If you produce 3D content, test a native-3D generator on one hero asset and check topology in your actual rig — speed means nothing if the mesh won't deform.
- Use the pay-on-export model to your advantage: generate many candidates free, download only the keepers, and budget credits for the final bake.
- For production adoption, verify commercial-license terms per tier; free tiers are typically non-commercial, and enterprise rights usually sit behind a paid plan.