VAST on Tripo P2.0 and Project Eden: Why 3D Generation in the Agent Era Is Still Underestimated

In August 2026, OpenAI co-founder and president Greg Brockman used X to repost a 3D interactive web project themed "Human Organ Teaching."

The developer behind it, @thebuggeddev, combined the modeling capabilities of the AI 3D large model Tripo P1.0 with the coding capabilities of GPT 5.6 Sol to quickly polish a beautifully crafted interactive web page—before this, he knew nothing about professional 3D modeling tools such as Blender, nor had he systematically studied programming. After the work was published, views on X quickly surpassed 2.1 million.

This was a highlight moment in @thebuggeddev's creative career. At the same time, a group of "AI + 3D" native creators also began to emerge.

Creators Haixin and Awen used Nano Banana to generate reference images, called Tripo P2.0 to complete 3D modeling, and then handed it to GPT-6 Astra to orchestrate scenes, special effects, and interaction logic, turning the Alice's Adventures in Wonderland chapter "A Conversation with the Caterpillar" into an exquisite interactive game; the production tutorial even made it onto Bilibili's trending list;

Independent developer Sac used a workflow of Claude Opus 5.5 + Tripo + Blender to turn every planet in The Little Prince into an explorable 3D scene in just 16 days; the work received tens of thousands of likes on X and was reposted by tech institutions and KOLs such as a16z and Marc Lou.

3D creation is beginning to become a new form of expression, even a new way of life. It no longer belongs only to professional modelers, game developers, and animators. Once a person has an idea, they can try to build their own world; a scene that originally existed only in the mind also has the chance to become a work that can be entered and interacted with.

And the main force driving this change is a group of young people trying to redefine the boundaries of 3D creation.

The core developer of the Tripo P series models used by @thebuggeddev, Haixin, and Awen is Ronnie, who just graduated with a PhD from Tsinghua University;

Project Eden, as the industry's first world model to propose a layered route of "Coding Agent + real-time generative renderer," has as its core researcher Zion, a master's student born in 2002.

Among them, some participated in core model training before even graduating and becoming full-time employees, while others started from a seemingly marginal interest and gradually pushed forward a new product and technology route.

These young people are all gathered at VAST (Sanqi Wanwu), a general artificial intelligence company founded in 2023.

Three years ago, VAST started in an era when AI 3D exploration was almost barren; today, tens of millions of users worldwide express themselves through its AI 3D model Tripo. From making models generate 3D assets faster and better, to enabling Agents to understand, invoke, and edit 3D content, and further to exploring how to build worlds that can run continuously and interact in real time, VAST is extending the reach of its research into broader spaces.

Team member R compared this process to "scientific research in Antarctica": along the way, everyone may fall into Antarctic ice crevices, or face Antarctic beasts, but the most important thing is to first "plant a ladder" and "build an elite exploration team"—in the end, the team will feel that "Antarctica isn't that cold after all."

01

Why do we need to reunderstand 3D?

For a long time, 3D creation has been tied to "professional, niche, high barrier to entry." In the public imagination, this creative process, which requires a long learning cycle and complex software, belongs to a small number of professional practitioners such as game developers, animators, and industrial designers.

The market's response is also indirectly reflected in the tech industry's overall investment in 3D. Before graduating, Ronnie interned at many major companies. But he found that for an industry like 3D, which is younger than LLMs and reaches practical application later, major companies were not willing to invest as many resources as expected.

R had a feeling from a different angle. Before joining VAST, he had deeply participated in cooperation with several major companies. But he found that being a major company does not necessarily mean having a winner mindset: "Much of the research at startups is more cutting-edge than at major companies, because everyone holds the courage to define problems themselves and the determination to win."

The people who joined VAST all adhered to a belief that was not mainstream at the time: 3D is the "mother tongue" of the world, the native expression of all things in digital space. The shape, structure, and position of objects, as well as the spatial relationships among them, can all be expressed through 3D.

Zion said bluntly: "3D is underestimated." He gave several examples: in the currently closely watched code world model route, what the upstream engine maintains is precisely the evolution of 3D content. Similarly, in the simulation training environments of embodied intelligence, 3D content will still remain the foundation for building environments for a long time to come. In the future, 3D will not only serve games, film and television, and industrial design, but also has the chance to become the foundation for AI to understand environments, operate objects, and build interactive worlds.

In just two years, this judgment gradually settled into Tripo's two model routes: the P series is committed to making generated 3D assets have production-grade topology and usability, while the H series continuously pushes the boundaries of high-precision geometry generation.

Since 2026, as LLMs' understanding, planning, and coding capabilities have been strengthened as never before, the industry has begun to reassess the barriers to 3D creation and its prospects for real-world deployment—when the upstream Agent is powerful enough, it can take over more and more 3D professional tools and creative workflows on behalf of ordinary people.

However, although Coding Agents can already generate Three.js pages, write some interaction and animation logic, and continuously debug based on runtime results, from "the Agent can write code" to "the Agent can develop interactive 3D applications," one key link is still missing: 3D assets that the Agent can quickly understand, edit, and drive.

Filling this gap is not easy. On the one hand, bringing 3D into more diverse creative workflows and tasks means that AI-generated 3D assets must be "usable as generated," or even genuinely good; on the other hand, as AI Agents take over more creative workflows, assets created by AI 3D models also need to be directly editable and riggable by AI, and able to enter real-time environments.

Today, Tripo's evolution and the development of Agents form a clever intertextuality. Tripo's H series models continuously improve the fidelity of complex forms and details; Tripo's P series, meanwhile, directly generates production-grade meshes in seconds—among them, P2.0, released in August 2026, supports native quad topology, has more regular edge flow, and supports natural part separation that better matches structural semantics, with unprecedented pipeline compatibility. These features also make P2.0 naturally suited to being understood and driven by Agents.

"Tripo's technical philosophy is 'production-ready.'" VAST Chief Scientist Cao Yanpei has said this many times. He is both the developer of P2.0 and one of its first creators. He had P2.0 generate a 3D mechanical spider made up of individual parts, then had Claude Opus write the motion logic, successfully making the spider move.

Just two weeks after P2.0's release, GPT-6 Astra dropped like a bombshell—this model, billed as "the world's most intelligent," ushered AI 3D into its strongest promotional period yet.

Overseas creator Nano added dynamic facial expressions to an anime-style character generated by Tripo based on Astra, earning 1,400 likes on X; @luccacerf, who has no modeling or game development experience, used the Astra-Tripo P2.0-Blender-Unity workflow to create a World of Warcraft fan game with just 3 prompts and a 90-minute production cycle—within just one week, Tripo API calls grew rapidly, and hundreds of members spontaneously formed an Astra+3D creation group inside VAST.

△Creator Nano (@Dstudio_ai) added dynamic facial expressions to an anime-style character generated by Tripo based on Astra.

This also gave new real-world substantiation to the judgment VAST has held since its founding: 3D is not just a category of content production; as Fei-Fei Li has judged, it will also become "a universal interface between humans and machines."

As Agents gradually become the new creators, the boundaries of 3D have only just begun to open.

March 2026, San Francisco, USA. At GDC, the world's top game developers conference, a demo of an AI 3D model called "Tripo P" drew the attention of many attendees. An overseas game creator praised Ronnie, a core developer of the P series, repeatedly: "This is absolutely amazing!"

In the AI 3D field at that time, Tripo P was an astonishing "new species." Previously, AI-generated 3D assets not only took minutes to generate, but their topology and edge flow quality also fell far short of production requirements, often relying heavily on human rework. But with the AI 3D model Tripo P series, the traditional AI 3D "impossible triangle" of quality, speed, and pipeline usability disappeared.

For Ronnie, the P series story reads like a wish-fulfillment novel. In July, at SIGGRAPH 2026, the top international conference on computer graphics held in Los Angeles, USA, a real-time demo based on Tripo P1.0 won the top prize, Best in Show, in the Real-Time Live! session; five VAST papers were accepted to Technical Papers, including the paper on Nexus, the underlying algorithm of the Tripo P series. In August, the P series' latest achievement, P2.0, became the world's first AI 3D foundation model to achieve end-to-end generation of native quad topology meshes on a diffusion model.

Spotlights, honors, applause—behind all of this lies a lonely quest undertaken amid non-consensus.

Ronnie remembers that when the P series was first established as a project inside VAST, few in the industry talked about using diffusion models for mesh generation. At that time, the huge success of the autoregressive paradigm in large language models deeply influenced the 3D field. Many researchers tried to copy the serialization approach directly onto 3D representations—this route did quickly give the industry a taste of "sweetness": based on the autoregressive route, 3D models could generate native meshes with low polygon counts, compact structures, and a certain production value.

But Ronnie realized very early that the autoregressive route brought a kind of "false prosperity" to native mesh generation. He mentioned that forcibly cramming meshes into serialized representations causes information loss, and once one pursues a higher ceiling, problems such as error accumulation will erupt all at once.

The idea of achieving mesh generation based on diffusion models quickly gained internal support—even though this was a river no one had successfully crossed, and even though Ronnie was still an intern at the time. "We had many opportunities to try, fail, and grow," he recalled. "We could do some out-of-the-box exploration ourselves."

VAST's exploration of world models likewise began along an almost deserted path.

Zion remembers that when he first joined the Project Eden team, the mainstream world model approach on the market was still the interactive real-time video generation route led by Genie 3. But members realized that with an input method based purely on keyboard and mouse actions, it was very difficult to create content that users would be willing to keep playing and that had real commercial value. More importantly, although action-conditioned video models could generate film-grade visuals, the underlying state was difficult to maintain.

A new "layered" interaction paradigm soon emerged: an upstream Coding Agent and interaction engine maintain the world's state and rules; a downstream real-time video generation model renders high-fidelity visual presentation.

As the first to try it, the difficulty of the research can be imagined. At the time, there was not much mature open-source code to reference, and the research team experimented with multiple versions of configurations, spending nearly half a year before getting the results to work.

In June 2026, as soon as Project Eden was released, it attracted attention at home and abroad. Tech KOLs such as Nataniel Ruiz (a researcher at Google DeepMind) and Matt Workman (a creative technology expert at fal) reposted it one after another.

Afterward, more and more world model research with "layering" as its philosophy appeared in the industry: not long after Project Eden was released, Li Feifei's team proposed a world model framework composed of a planner, a simulator, and a renderer; Genie 4, released at the end of October, also pointed out in its blog that hybrid world models may be the key path for the next stage: integrating the structured reasoning of code and language models with the rich perceptual capabilities of pixel-level simulation.

Opportunities for technological innovation are often hidden where consensus has not yet arrived. From the Tripo P series to Project Eden, VAST has chosen paths that few people took at the time.

This boldness has also gradually translated into VAST's technological competitiveness. Since its founding, VAST has had more than 50 open-source projects and has cumulatively published more than 60 top-conference papers in the 3D field, making it one of the teams with the largest reserve of top-conference papers in the global 3D field. In blind tests conducted by an independent third-party evaluation institution based on more than 200,000 users, the overall performance of the Tripo series models has steadily ranked first over the past six months.

This achievement also makes a simple truth evident once again: truly valuable opportunities often belong to those who dare to walk ahead of consensus.

03

Next step: become a V·STAR

Walking ahead of consensus requires not only technical judgment, but also a group of people willing to bear uncertainty for that judgment.

Inside VAST, this group has been given a rather romantic title: V·STAR.

V·STAR is a dedicated program established by VAST for top young technical talent worldwide. The program offers uncapped salaries, options, and computing power and resources matched to research tasks. Currently, V·STAR is looking for young researchers in directions such as 3D generation, 3D understanding and agents, and world models. Whether one has previously worked deeply in computer vision, graphics, or large language model post-training, Coding Agents, or video generation, as long as one has one's own research and judgment on these issues, there is an opportunity to find new topics here. (Official application channel: https://tripo3d.com/vstar; application email: vstar.talent@vastai3d.com)

In VAST's view, top talent is not defined only by school and the number of papers. Compared with a polished resume, the team values more whether a person has put forward valuable judgments, whether they have led important work, and whether they have truly advanced a difficult problem.

In R's view, exploring unknown fields requires not a huge team, but an elite team capable of independently handling various problems. "When establishing a scientific expedition team, what you need is not a large army, but an elite team, where each member may be able to take on multiple roles. This is also why I highly recognize VAST, especially the Research team."

This lean organizational form also affects the team's way of working. R observed that startups often have more work than people, and everyone has the opportunity to directly take on important tasks; large companies may have more people than work, good projects need to be fought over, and internal friction is likely to arise between teams.

"Everyone says 'no upper limit,' but the real no upper limit is: whatever you want to do, you can get the opportunity to do it."

When R joined VAST, he wanted to try things he had not done before. After he communicated with the team, his ideas were quickly recognized. "In the end, once we did it, it worked."

This autonomy also means that researchers can more directly take responsibility and have a longer exploration cycle.

At VAST, Ronnie's experience is one of freedom. For research with long cycles and uncertain outcomes, this kind of space is especially important. He mentions that colleagues share a similar understanding of the value of research, and no one would demand that the team produce a certain result by next month. "Everyone respects the process of scientific research and allows failure to happen; when problems come up, we discuss solutions together."

Discussions within the team are also quite direct. When he first joined, R still worried about his position and whether expressing different opinions would overstep boundaries. But he soon discovered that almost every issue could be put on the table for discussion. In a recent exchange, several members came up with an experimental idea on the spot, formed a team right there, and quickly ran the experiment—and the results were a pleasant surprise.

This short distance from idea to action is also evident among interns. Stewie, a fifth-year PhD student, once fully trained an online core model at VAST. "At many companies, interns don't get to touch this," he said. "If you think of something, you can just go do it; there isn't a lot of cost wasted in between on other things, and there's no horse-racing mechanism."

Of course, a scientific expedition team also needs enough equipment and supplies. R's impression is that researchers at VAST can sometimes actually obtain more abundant resources than at big companies. Especially in research directions that do not rely heavily on compute, the team's organizational efficiency and way of using resources can create a lot of room to perform.

VAST also continues to provide funding support for research investment. Over the past six months, the company has completed a total of about 5 billion RMB in financing. For AI 3D and world model research, which is still in a rapid exploration phase, stable funding, compute, and talent investment are the foundation that allows a "scientific expedition team" to keep pushing into unknown territory.

So what kind of person is suited to join this team?

The first criterion R gives is: high capability, low ego.

He once again uses the Antarctic expedition team as a metaphor: in the ice and snow, team members first have to consider how to complete the mission; no one keeps arguing about who is the number one position. Applied to VAST, the team does not particularly care about which platform a person came from in the past, nor does it screen talent purely by the number of papers and academic credentials. What matters more is whether he has the courage to break the ice—when facing a problem that has never been solved, can he boldly put forward his own ideas based on his judgment of the essence of the technology and put them into practice?

In R's view, for truly capable researchers, platform endorsement is not that important: "Your value depends on the depth of your thinking about the essence of the problem and your research taste; when you get the opportunity to explore, this value will ultimately be reflected in your output."

At VAST, everyone has the opportunity to participate deeply in the key links of a project and personally drive a piece of research from idea to result. Compared with being in a huge organization and becoming a participant in some project, this high degree of participation also gives young researchers more opportunities to build their own technical influence.

In his campus recruitment year, Ronnie also asked himself: will I still be able to find a better opportunity to start a business in the future?

"I find it hard to imagine myself working at a big company for three or four years and then coming out to start a business." His answer to himself was: "Anyway, it's my first job, and there's not much to lose. Since this possibility exists, I might as well try it while I'm young."

Over the past three years, the model Ronnie participated in developing has already been integrated by Agents into creative workflows. From a technical judgment that was not yet widely recognized by the industry back then to a tool that more and more developers can now use, he has personally experienced the process of a non-consensus path gradually becoming reality.

At the end of the conversation, R also left a rather meaningful comment on his own "Antarctic trip": "That said, I feel Antarctica isn't actually that cold."

Of course, unknown places will not stay cold forever. Someone walks in first, someone sets up a ladder, and only then do those who come later have the chance to continue forward along this path.

As Agents begin to build worlds, how the next generation of AI will understand space, manipulate objects, and keep a world running still involves many uncharted territories. What VAST hopes to find are exactly those people who are not in a hurry to wait for answers and are willing to walk in first.

(Ronnie, Zion, R, and Stewie are all pseudonyms in this article.)