A game studio needs 200 background crates, each slightly different, ready to drop into a level by Friday. Outsourcing them means a week and a purchase order. A 3D-generation model turns that into an afternoon — if the output is clean enough to survive the engine. VAST, the company behind Tripo (三启万物), has spent two years arguing that the bottleneck was never imagination but geometry.
Its answer arrived in August 2025 as Tripo 3.0, a model the company says runs at 20 billion parameters (200亿) and is built to produce assets that need little human cleanup.
Key takeaways
- Tripo 3.0 (August 21, 2025) scales to 20 billion parameters (200亿) and uses SparseFlex, a geometry representation designed for higher resolution and efficiency.
- VAST open-sourced TripoSG — the first Mixture-of-Experts Transformer used in 3D generation (1.5B parameters) — and TripoSF, which can generate a model's interior structure, not just its exterior.
- The open models were trained on a 2-million image-SDF dataset the company built through a quality-controlled pipeline.
- Per official-disclosed ecosystem figures, VAST serves over 3 million global developers and 700+ enterprise clients, with more than 40 million high-quality 3D native assets generated.
The open-source bet
VAST's strategy has been unusually generous for a revenue-stage startup: give the research community the foundations, then monetize the workflow on top. In March 2025 it launched an "open-source month" and released TripoSG and TripoSF.
TripoSG matters because it ports the Mixture-of-Experts Transformer — common in language and image models — into 3D for the first time, the company says. The 1.5B-parameter release produces quality on par with Tripo 2.0 and, VAST claims, surpasses all existing open-source 3D generation. TripoSF, built on the SparseFlex representation, goes further: it can generate not just the back of an object but its interior — the seat frame of a bus, the cabin of a driver's seat — and handles open surfaces like clothing or petals without the bloated geometry earlier methods produced.
The training data is the quiet differentiator. VAST says it holds one of the largest collections of high-quality native 3D data and runs it through a pipeline of quality scoring, filtering, fixing and SDF production, yielding 2 million clean image-SDF pairs.
From 360p to IMAX, in 3D terms
Founder and CEO Song Yachen (宋亚宸) has repeatedly framed 3D progress in a camera-resolution analogy: Tripo 1.0 (2024) was roughly "360p," Tripo 2.0 (September 2024) "720p," and the gap between models would be decided by resolution, not by product polish. Tripo 3.0, he argues, is the version that finally reaches "IMAX" — topology, control and detail high enough to enter most production pipelines.
Concretely, Tripo 3.0 adds pose control, symmetry control, multi-view image input and smooth-detail control, letting a user sketch a rough shape and get a refined 3D model in seconds, auto-rig a T-pose character, or paint materials with a smart texturing brush. The company reports partnerships spanning Tencent, NetEase, Bambu Lab and HTC, and integrations with Unity and Blender.
Why this matters beyond gaming
3D generation has long been the quiet cousin of text, image and video models — harder to judge and harder to use. VAST's open-plus-commercial dual track is an attempt to make it infrastructure: a foundation other tools build on, with the company earning through the Tripo Studio workflow and enterprise API rather than a closed app.
Reported financing underscores the bet. Industry media described a 2026 strategic round exceeding RMB 1 billion (≈ US$140M / HK$1.1B) for VAST, alongside earlier rounds — a signal that game and hardware firms are willing to fund 3D generation as core infrastructure rather than a curiosity.
From a US$20 subscription to a unicorn
The open-weights story is only half of VAST's bet; the other half is a paying user base. The company charges from US$20 a month plus token-based fees, and founder Simon Song (宋亚宸) told Bloomberg that most of its roughly 20 million users sit in the United States, with Europe, Japan and South Korea next. That overseas skew mirrors Kling's: Western studios pay for asset pipelines that Chinese rivals often give away. In June 2026 VAST closed a near-US$200 million round led by INCE Capital and a China Life-backed fund, pushing its valuation past US$1 billion and making it one of China's newer AI unicorns. A concrete proof point is the NetEase (网易) game Where Winds Meet (燕云十六声), where players can turn an uploaded photo into a 3D avatar on the spot — generation that has already reached players' hands, not a roadmap slide. VAST says its P-series model now outputs low-poly, game-ready meshes to skip the hours of manual cleanup that have kept AI 3D out of shipping titles.
What readers can do now
- If you build games or XR, prototype with TripoSG/TripoSF — the open weights let you test asset generation without committing to a commercial tier.
- Evaluate Tripo 3.0 on your worst-case asset (thin fabrics, interiors, multi-part objects) where earlier 3D models degrade, and measure cleanup time saved.
- Plug into your existing toolchain via Unity or Blender integrations before assuming a standalone workflow; the production value is in the handoff, not the generation.
Honest limitations
- Tripo 3.0's 20-billion-parameter figure and the "first MoE Transformer in 3D" claim come from VAST and corroborating tech-media reports, not an independent architecture audit I could verify.
- Ecosystem numbers (3M+ developers, 700+ enterprise clients, 40M+ assets) are officially disclosed via media; I did not find a third-party confirmation.
- The RMB 1 billion+ financing round is reported by industry media, not a primary regulatory filing I could access; treat it as reported.
- "Industrial-grade" is VAST's positioning; real production fit depends on your engine, asset standards and tolerance for manual cleanup, which varies widely.