Nila Kartika Wati
In an era witnessing the rapid evolution of artificial intelligence, a crucial shift is underway, moving AI’s capabilities beyond mere content generation towards the creation and simulation of interactive environments. This paradigm shift, centered on "world models," marks a significant frontier in AI research and commercialization. Amidst this global technological race, Chinese startup HiDream.ai has emerged as a formidable player, recently making waves with its innovative HiDream-O1-World model. This development underscores the company’s strategic pivot and its ambition to carve out a distinct niche in the competitive landscape of generative AI, particularly within China’s vibrant and often underestimated tech ecosystem.
The heightened interest in world models received substantial validation this year when Fei-Fei Li’s World Labs, a high-profile AI pioneer, successfully raised a staggering $1 billion in funding and concurrently launched Atlas, an omni world model designed to span text, images, video, and 3D. This monumental investment signals a clear industry consensus on the strategic importance and immense potential of AI systems capable of understanding, simulating, and interacting with complex virtual worlds. Against this backdrop of significant global activity, HiDream.ai’s launch of HiDream-O1-World on August 16 represented a bold move beyond its foundational strengths in image and video generation, positioning the company directly at the forefront of interactive AI development.
HiDream-O1-World distinguishes itself by accepting diverse inputs—text, images, and interactive commands—allowing users to not only generate environments but also to actively navigate these simulated spaces, altering characters, objects, and conditions dynamically as scenes unfold. Upon its release, the model achieved a remarkable score of 80.9 on WBench’s Navi leaderboard, an independent benchmark specifically designed to evaluate the navigation and interaction capabilities of world models. This achievement initially placed HiDream-O1-World at the top of the rankings, signaling a significant technical breakthrough. While newer models have since surpassed its initial position, by September 15, it maintained its score of 80.9, holding a respectable No. 6 ranking, a testament to its robust capabilities in a rapidly advancing field.
Architects of Innovation: Leadership and Vision
HiDream.ai was founded in March 2023, driven by a singular vision: to build cutting-edge generative multimodal models. The company’s genesis is deeply rooted in the distinguished academic and industrial background of its founder and CEO, Tao Mei. Mei’s career trajectory is a testament to his expertise in computer vision and multimodal AI. He spent 12 impactful years at Microsoft Research Asia, a renowned hub for AI innovation, where he honed his skills and contributed to foundational research. Following his tenure at Microsoft, Mei transitioned to JD.com, one of China’s e-commerce giants, where he initially led the company’s computer-vision research efforts before ascending to the role of vice president. His extensive work consistently combined vision and language processing, exemplified by his influential 2017 ACM Multimedia paper, "To Create What You Tell: Generating Videos from Captions," which explored the then-nascent field of generating dynamic visual content from textual descriptions.
The intellectual capital within HiDream.ai is exceptionally high, reflecting Mei’s commitment to scientific rigor. Over 90% of the company’s core technical staff hold either master’s or doctoral degrees, largely from prestigious institutions, indicating a strong emphasis on deep research and development capabilities. This academic pedigree provides a solid foundation for tackling complex AI challenges. The company has also experienced rapid growth in its workforce, expanding from fewer than 50 employees in early 2025 to an estimated 200 to 300 by May 2026. This aggressive hiring strategy underscores the ambitious scope of its projects and the significant investment in human capital required to compete in the high-stakes AI sector.
Mei’s strategic decisions in the early days of HiDream.ai were critical. After departing JD.com, he contemplated various AI avenues, including robotics and large language models (LLMs). However, he deliberately chose to avoid a direct confrontation in the increasingly resource-intensive LLM race, which by then demanded thousands of expensive accelerators for competitive development. While image and video generation also required substantial computational resources, Mei perceived the scale to be "more manageable" and, critically, saw a clearer and more immediate path to commercialization. This pragmatic approach allowed HiDream.ai to focus its limited resources on areas where it could achieve differentiated technical and market advantages, setting the stage for its subsequent innovations.
From Pixels to Worlds: A Journey in Generative AI
HiDream.ai’s journey began with securing its initial external funding, notably from a syndicate of 15 alumni of the University of Science and Technology of China, highlighting the strong network effects within China’s scientific community. By December 2023, the company had successfully completed two funding rounds, cumulatively raising close to RMB 100 million. These early investments fueled the development of its foundational products, including Pixeling, a visual-creation platform catering to general users, and PixMaker, tailored for the specific demands of e-commerce imagery.
A cornerstone of HiDream.ai’s early technical strategy was its "dual-model" approach, which involved developing image and video models in parallel. This methodology allowed the company to leverage the comparatively cheaper image model as a testing ground for new ideas and architectural innovations before porting successful concepts to the more computationally intensive video domain. Mei publicly stated that this strategy significantly reduced training costs, estimating them to be roughly one-fifth of the industry average—a critical advantage for a startup competing with well-funded tech giants.
However, the path was not without its challenges. Mei candidly admitted that HiDream.ai’s first video model, released in August 2023, was "terrible." This self-assessment reflects the iterative and often difficult nature of AI development. The competitive landscape shifted dramatically with OpenAI’s unveiling of Sora, a highly capable video generation model, which prompted Mei’s team to re-evaluate their architectural choices. In response, HiDream.ai pivoted towards a Diffusion Transformer architecture and began exploring hybrid autoregressive approaches, adapting their research direction to incorporate the latest advancements and maintain competitiveness.
This evolutionary path culminated in the development of HiDream-O1-World. The model is built upon HiDream.ai’s proprietary Unified Transformer architecture, designed to process and integrate diverse data modalities—text, images, video, and crucial spatial information—within a single, coherent framework. A key technical innovation within this architecture is its ability to separate scene geometry from visual appearance. This separation is vital for preserving spatial consistency and realism during user navigation and interaction within generated environments, addressing a common challenge in interactive world models. For a startup operating with fewer resources than established technology giants, such deliberate and innovative architectural choices are paramount for optimizing compute and engineering investments, ensuring maximum impact from every dollar spent.
Refining the Product-Market Fit: Beyond Model Capability
Perhaps the most profound lesson HiDream.ai learned came directly from its engagement with customers. The company initially operated under the assumption that superior model capability would inherently translate into a compelling product. This "model is the product" mindset, common among research-driven startups, proved insufficient for the demands of the enterprise market. Business customers, as Mei later recounted in a May 2026 interview with 36Kr, were not looking for raw AI models that they had to integrate and assemble themselves. Instead, they sought comprehensive software solutions that could efficiently complete specific jobs and integrate seamlessly into their existing workflows.
"We initially thought model capability was the product," Mei confessed, highlighting a critical turning point in the company’s strategic thinking. HiDream.ai eventually recognized the imperative of developing an "agent or workflow layer" to bridge the gap between its powerful foundation models and the practical needs of business users. This realization spurred a significant re-evaluation of its product strategy.
The e-commerce sector proved to be an invaluable proving ground for this refined approach. Initially, HiDream.ai targeted conventional product imagery, but quickly discovered that such assets often remain unchanged for extended periods, leading to limited recurring demand. This insight prompted a strategic shift towards content-driven commerce and marketing. In this domain, brands frequently require thousands of short videos per month for dynamic campaigns, product launches, and social media engagement. Here, AI could transition from a standalone tool to an integral part of a continuous content production workflow, generating consistent, high-value demand.
This strategic pivot also necessitated a rethinking of HiDream.ai’s monetization models. The company experimented with various revenue strategies, including subscription services for its platforms, content creation services, and even, in specific marketing campaigns, revenue sharing models tied to gross merchandise value (GMV). These practical experiments and customer feedback informed the development of its broader "1+1+3" strategy. This comprehensive framework integrates three core components: powerful foundation models at its base, an enterprise model-and-agent platform that acts as the crucial intermediary layer, and specialized applications tailored for high-demand sectors like marketing, film and television production, and social media content creation. Critically, the platform is designed with flexibility, capable of leveraging third-party models when they offer a better fit for a particular task, ensuring optimal performance and utility for its diverse clientele.
The Economics of Advanced AI: Efficiency and Competition
For a startup like HiDream.ai, operating with finite resources compared to the vast budgets of technology giants, achieving high "performance per dollar" is not just an advantage—it’s an existential necessity. This focus on efficiency has been a consistent theme throughout the company’s development.
Mei elaborated in a 2025 interview with The Paper that HiDream.ai’s innovative image-video development loop, a key component of its dual-model strategy, significantly reduced training costs. Furthermore, the company projected that ongoing architectural enhancements would lead to a reduction of more than half in video inference costs. These cost-saving measures are critical in an industry where computational expenses are a major barrier to entry and scaling. In June, Mei further highlighted the commercial impact of HiDream.ai’s tools to Xinhua, stating that they could reduce the cost of producing a one-minute commercial short video to roughly one-tenth of traditional production methods. This dramatic cost reduction positions HiDream.ai as a powerful enabler for businesses seeking to scale their video content creation without prohibitive expenses.
The intense competition in the AI landscape is starkly illustrated by the performance of HiDream.ai’s models on public benchmarks. Its HiDream-O1-Image-1.5 model achieved a top-three ranking on Artificial Analysis’s text-to-image leaderboard in June, notably outperforming Google’s Nano Banana 2 at the time. However, the transient nature of technical leads in AI became evident quickly; by September 2, the same model had shifted to the No. 16 position. Artificial Analysis also listed its API cost at approximately $80 per 1,000 images, providing transparency on its commercial offering. This rapid fluctuation underscores the blistering pace of innovation in generative AI, where even leading models can be surpassed within weeks or months by new advancements.
Recognizing the value of developer engagement and ecosystem building, HiDream.ai has also strategically embraced open source. In May, the company released its 8-billion-parameter HiDream-O1-Image model and its accompanying code under an MIT license. This move, a departure from a traditionally more closed approach, was influenced by the success of other open-source initiatives in AI. Mei has acknowledged that DeepSeek, a prominent Chinese AI company known for its open-source contributions, "changed his thinking" about the benefits of sharing models, even when the immediate commercial return might not be clear. This open-source strategy aims to foster a broader developer community, accelerate innovation, and establish HiDream.ai’s models as industry standards, ultimately enhancing its long-term strategic position.
Securing the Future: Funding and Investor Confidence
In July, HiDream.ai announced a significant milestone: a RMB 1.5 billion Series C funding round. This substantial investment brought its total financing over the preceding three months to more than RMB 2.1 billion, reflecting strong investor confidence in its technological capabilities and business model. Such a considerable capital injection provides HiDream.ai with the necessary resources to further accelerate its research, expand its operations, and scale its enterprise solutions.
Investors are keenly scrutinizing AI startups for specific attributes that signal long-term viability and competitive advantage. Wang Bing, a representative from Oriental Fortune Capital, an investor in HiDream.ai, articulated his firm’s criteria. They sought companies capable of building "competitive foundation models at lower cost," demonstrating "high R&D and capital efficiency," and exhibiting "fast translation into enterprise use cases." These criteria align perfectly with HiDream.ai’s strategic focus on cost-efficient innovation and direct commercial application.
A particularly crucial factor highlighted by Wang Bing was HiDream.ai’s commitment to "legally licensed visual data." In an AI landscape increasingly fraught with copyright concerns and legal challenges surrounding training data, the company’s proactive approach in accumulating 200,000 hours of licensed video through partnerships provides a significant competitive moat and mitigates substantial legal risks. This emphasis on ethical and legally sound data practices not only protects the company but also enhances its appeal to enterprise clients who demand compliance and reliability.
Despite these successes, the economics of advanced AI remain demanding. Mei candidly acknowledged to Xinhua that "talent, data, and compute remain expensive." He projected that HiDream.ai anticipates reaching monthly break-even only by 2029, illustrating the long investment horizon and significant capital expenditure required to develop and scale cutting-edge AI technologies. This long-term financial outlook underscores the scale of ambition and the sustained commitment necessary for HiDream.ai to achieve its goals.
Strategic Divergence: HiDream.ai’s Enterprise-First Path
To fully appreciate HiDream.ai’s strategic trajectory, it is useful to contrast its approach with that of global technology giants like Google DeepMind. Google’s visual-AI portfolio is expansive, encompassing models like Nano Banana and Imagen for images, Veo for video, and Genie 3 for interactive world models. These offerings are seamlessly integrated into Google’s vast existing consumer, developer, and cloud ecosystems, providing inherent distribution advantages and brand recognition.
HiDream.ai, lacking comparable built-in distribution channels and an established consumer base, has adopted a fundamentally different route. From its inception, the company has treated enterprise services as a core, rather than an ancillary, part of its strategy. This focus has shaped every aspect of its development, from model architecture to business model innovation. The central challenge for HiDream.ai has been to effectively serve businesses through tailored solutions, whether via software subscriptions, specialized content services, or models whose performance is directly tied to tangible customer results.
This enterprise-centric focus is vividly reflected in its integrated model-platform-application stack and its strategic shift from low-frequency product imagery to addressing the recurring, high-volume demand for content in marketing and e-commerce. The extension of its technical capabilities into interactive world models with HiDream-O1-World further broadens its appeal to industries requiring sophisticated simulation and virtual environment creation. By the first quarter of 2026, HiDream.ai’s products had garnered an impressive user base, serving over 30 million professional users and more than 40,000 enterprise customers worldwide. This global reach, achieved through a targeted enterprise strategy, highlights the effectiveness of its specialized approach in a highly competitive market.
A Scientist’s Business Odyssey: Evolving Leadership and Brand
HiDream.ai’s organizational culture and strategic direction still bear the strong imprint of Tao Mei’s research background. In a 2021 JD profile, Mei articulated a philosophy common among deep-tech scientists: he argued that scientists should be afforded the freedom to pursue work that is "the best" or "the first," prioritizing fundamental breakthroughs, while engineers should focus on translating these innovations into standardized products and services. This delineation reflects a traditional academic approach, where the pursuit of knowledge often takes precedence over immediate commercialization.
However, by 2026, Mei demonstrated a pragmatic evolution in his leadership perspective, acknowledging the need for HiDream.ai to move beyond its traditionally low-profile approach. He recognized the imperative to "strengthen its brand narrative," a crucial aspect of gaining market visibility and attracting both talent and customers in a crowded AI space. "We come from a scientist-entrepreneur background and are used to keeping our heads down and doing the work," he remarked, indicating a conscious effort to balance deep scientific pursuit with the demands of building a public-facing, commercially successful company. This shift in mindset from a purely scientific endeavor to a market-driven enterprise is a common yet critical transition for many founder-led deep-tech startups.
Looking Ahead: The Road to Sustainable Growth
HiDream.ai stands at a pivotal juncture. The next stage of its evolution will be defined by its ability to effectively translate its advanced research, sophisticated enterprise workflows, and ambitious expansion into world models into a durable and sustainable business. Its success will hinge on its capacity to not only maintain its technical edge in a fiercely competitive environment but also to continuously refine its product-market fit, effectively communicate its value proposition, and navigate the complex economic realities of the AI industry.
The company’s journey from a research-intensive startup focused on image and video generation to a leader in interactive world models, underpinned by a clear enterprise strategy and a commitment to cost efficiency, offers a compelling case study in China’s rapidly maturing AI ecosystem. As the global AI landscape continues to evolve, HiDream.ai’s trajectory will undoubtedly provide valuable insights into the challenges and opportunities for innovative startups seeking to make a lasting impact on the future of artificial intelligence. Its strategic choices, from architectural design to business model adaptation and brand building, will shape its path towards achieving its ambitious vision and contributing significantly to the next wave of global innovation.
Evan Lee Salim
