Editor’s Note: There is no shortage of technology news, yet many promising Chinese startups still remain largely unknown to global audiences. At the same time, the constant flow of funding news, product launches, and new technologies makes it harder to see which companies truly deserve attention. DragonLab is TechNode’s effort to close that gap, looking past the noise to discover China’s next wave of innovation—emerging companies, ambitious founders, and breakthrough technologies that have yet to receive global attention.
Interest in world models is rising as AI moves from generating content toward simulating interactive environments. Fei-Fei Li’s World Labs underscored that shift this year with a $1 billion raise and the launch of Atlas, an omni world model spanning text, images, video and 3D.
When HiDream.ai launched HiDream-O1-World on Aug. 16, it moved beyond image and video generation into interactive world models. The model accepts text, image and interactive inputs, letting users navigate generated environments and alter characters, objects and conditions as scenes unfold. It scored 80.9 on WBench’s Navi leaderboard and ranked first at the time. Newer models later moved ahead; as of Sept. 15, it remained at 80.9 but ranked No. 6.
HiDream.ai was founded in March 2023 to build generative multimodal models. Founder and CEO Tao Mei spent 12 years at Microsoft Research Asia before joining JD.com, where he led computer-vision research and later became a vice president. His work had long combined vision and language, including the 2017 ACM Multimedia paper To Create What You Tell: Generating Videos from Captions. More than 90% of HiDream.ai’s core technical staff hold master’s or doctoral degrees. The company had fewer than 50 employees in early 2025 and roughly 200 to 300 by May 2026.
After leaving JD, Mei considered robotics and large models but avoided a head-on LLM race that increasingly required thousands of accelerators. Image and video generation still demanded substantial compute, but the scale appeared more manageable and commercialization closer.
Starting with pictures to get to video
HiDream.ai’s first outside funding included 15 alumni of the University of Science and Technology of China. By December 2023, it had completed two rounds totaling close to RMB 100 million. Early products included Pixeling, a visual-creation platform, and PixMaker for e-commerce imagery.
The company developed image and video models in parallel, using the cheaper image model to test ideas before carrying them into video. Mei told The Paper this “dual-model” approach reduced training costs to roughly one-fifth of what he considered the industry average.
The early results were uneven. Mei later called HiDream.ai’s first video model, released in August 2023, “terrible.” After OpenAI unveiled Sora, his team shifted toward a Diffusion Transformer architecture and explored hybrid autoregressive approaches.
That path now extends to HiDream-O1-World, which uses HiDream.ai’s Unified Transformer architecture to process text, images, video and spatial information within a shared framework. It separates scene geometry from visual appearance to help preserve spatial consistency during navigation and interaction.
For HiDream.ai, smaller scale meant making more deliberate choices about where to commit compute and engineering resources.
When “the model is the product” stopped working
The harder lesson came from customers. HiDream.ai initially assumed better model capability would itself make a compelling product. But enterprise customers wanted software that could complete a job, not models they had to assemble themselves.
“We initially thought model capability was the product,” Mei recalled in a May 2026 interview with 36Kr. The company eventually realized it needed an agent or workflow layer between foundation models and business users.
E-commerce helped clarify what that meant in practice. HiDream.ai first targeted conventional product imagery, but those assets may remain unchanged for months, limiting recurring demand. The company then shifted toward content-driven commerce and marketing, where brands may need thousands of short videos a month and AI can become part of a continuous production workflow.
That shift also pushed HiDream.ai to rethink how it charged customers. It experimented with subscriptions, content services and, in some campaigns, revenue sharing tied to gross merchandise value. Over time, those experiments fed into its broader “1+1+3” strategy, which combines foundation models, an enterprise model-and-agent platform, and applications for marketing, film and television, and social media. The platform can also use third-party models when they better fit a task.
Performance, price and the problem with being No. 1
For a startup with fewer resources than technology giants, performance per dollar matters alongside absolute performance.
Mei told The Paper in 2025 that HiDream.ai’s image-video development loop reduced training costs, while architectural changes were expected to cut video inference costs by more than half. In June, he told Xinhua that HiDream.ai’s tools could reduce the cost of producing a one-minute commercial short video to roughly one-tenth of traditional production.
HiDream-O1-Image-1.5 reached the top three on Artificial Analysis’ text-to-image leaderboard in June, ahead of Google’s Nano Banana 2 at the time. By Sept. 2, it ranked No. 16. Artificial Analysis listed its API cost at about $80 per 1,000 images.
The movement shows how quickly technical leads can narrow. HiDream.ai has also used open source to build developer attention. In May it released the 8-billion-parameter HiDream-O1-Image model and code under an MIT license. Mei has said DeepSeek changed his thinking about open source, even when the immediate commercial return is unclear.
What investors are actually buying
In July, HiDream.ai announced a RMB 1.5 billion Series C round, bringing its financing over the previous three months to more than RMB 2.1 billion.
Investor Wang Bing of Oriental Fortune Capital said his firm looked for competitive foundation models built at lower cost, high R&D and capital efficiency, and fast translation into enterprise use cases. He also pointed to HiDream.ai’s legally licensed visual data as important in a field where copyright remains a significant risk. The company has accumulated 200,000 hours of licensed video through partnerships.
The economics remain demanding. Mei told Xinhua that talent, data and compute remain expensive and that HiDream.ai expects monthly break-even in 2029.
Building from the enterprise out
Google DeepMind offers a broader contrast to HiDream.ai. Its visual-AI portfolio spans Nano Banana and Imagen for images, Veo for video and Genie 3 for interactive world models, supported by Google’s existing consumer, developer and cloud ecosystem.
HiDream.ai has taken a different route. Without comparable distribution channels, it has treated enterprise services as a core part of its strategy from the start. The challenge was how to serve businesses through software subscriptions, content services or models tied more directly to customer results.
That focus is reflected in its model-platform-application stack and its move from low-frequency product imagery toward recurring content demand. HiDream-O1-World extends that technical range into world models. As of the first quarter of 2026, HiDream.ai’s products served more than 30 million professional users and more than 40,000 enterprise customers worldwide.
A scientist-led company learning to tell a business story
HiDream.ai still carries Mei’s research background. In a 2021 JD profile, he argued that scientists should have freedom to pursue work that was “the best” or “the first,” while engineers focused on standardized products and services.
By 2026, he was also reconsidering HiDream.ai’s low-profile approach, including the need to strengthen its brand narrative. “We come from a scientist-entrepreneur background and are used to keeping our heads down and doing the work,” he said.
For HiDream.ai, the next stage will be defined by how well it can translate research, enterprise workflows and its expansion into world models into a durable business.
Key milestones in the company’s evolution
Have a startup we should know about? Reach us at dragonlab@technode.com.