By Cheng Qian, reporting from Aofeisi (凹非寺)
Quantum Bit | Official Account QbitAI
Recently, the 2026 World Laureates Forum's Second International Conference on Engineering Intelligence was held. As an annual event of great influence in the global scientific community, this conference brought together top scientists, academicians, and industry leaders from around the world, with "Engineering Intelligence" as the core theme, to discuss frontier trends in the intelligent era together.
At the meeting,Mei Tao, a foreign member of the Canadian Academy of Engineering and founder and CEO of HiDream.ai (智象元一)delivered a speech, proposingBuilding Agentic Omnimodal World Models, toward the next generation of world intelligence.
AI model aggregation platformOpenRouter data shows that the scale of global large model calls is entering an explosive phase. As of October (10) this year, China's weekly Token call volume reached 93 trillion, 4.16 times that of the United States, ranking for 26 consecutive weeksfirst in the world。
Behind these numbers, AI is rapidly evolving from a “tool” into a new type of social infrastructure, like water and electricity.
At the same time, model iteration is accelerating constantly. In 2023, large models iterated on a roughly 12-month cycle. As Recursive Self-Improvement (RSI) pushes AI into a self-accelerating orbit, model iteration now advances on a weekly basis.
The stronger AI's capabilities become, the more its shortcomings cannot be ignored. In thefield of Digital AI, from text interaction and code writing to video generation and even mathematical research, AI has reached or even surpassed the level of human experts. Take mathematical research as an example: recently OpenAI released 377 mathematical results at once, covering hundreds of long-standing open problems.
The development of Physical AI, by contrast, has been relatively slow, constrained by weak physical understanding, scarce interaction data, and insufficient closed-loop execution capabilities.
How to bridge the gap between Digital AI and Physical AI is the defining question of our era that the next generation of artificial intelligence must answer.And world models are precisely the bridge across this gap.
The core of a world model is to enable AI to understand how the world works and to infer events that may happen in the future. Through unified understanding and generation of omnimodal information, it models world states, physical laws, and causal relationships, thereby achieving perception, inference, and prediction of the real world.
Different technical paths lead to world models, and this has become the next competitive focal point in AI development.The industry has now formed three evolutionary routes: large language models, multimodal models, and embodied large models. Mei Tao believes that these three technical routes are not isolated from one another; in the long run they will eventually converge towardAgentic Omnimodal World Models。
. Agentic Omnimodal World Models mainly consist of three parts:
First, theLLM，, which is the decision-making brain of the agent, responsible for understanding intent, generating reasoning, and formulating task plans.
Second, the Omni-Model,That is, a general world model. It is a world simulator: based on LLM intent, it simulates the operation of the physical world, deduces the consequences of actions, and predicts the next move.
Third, Embodied Context,Embodied Context. It connects to the real physical environment, executes embodied actions, and simultaneously feeds real observational data back to calibrate the world model.
The three work in synergy to achieve"One Token Space, One World"—in a single Token space, the world of thought and the physical world achieve unification. And this is precisely the key to the next generation of world intelligence.
His reflections on world models have been internalized into HiDream's product philosophy and technological DNA. HiDream independently developed anative omni-modal (Unified Transformer) model architecture, enabling unified representation across all modalities—text, images, video, space, pose, force and touch, sensors, and more. Based on the UiT foundation, HiDream has built a complete world model matrix: starting from an image model, extending to a video model, an interactive model HiDream-O1-World, and further to the embodied model HiDream-O1-Embodied, completing an evolution of capabilities from static images and dynamic video to spatial interaction and embodied action.
HiDream has already formed a "1+1+3" commercialization architecture: at the bottom layer is a native omni-modal world model matrix; in the middle, through the enterprise-level model platform HiHarness, it aggregates its self-developed HiDream-O1 omni-modal series of models as well as the world's top models; at the top layer are three agents, including the AIGC commercial marketing agentHiBurst, the film and video creation agent"Zhenzan", and the content creation agentvivago。
At the same time,HiDream is also vigorously promoting the practice of world models in frontier fields. In the field of embodied intelligence, the world model can synthesize millimeter-level, high-precision action videos at scale, breaking through the bottleneck of high cost and small scale of real-machine data, providing data supply for real-machine training; in the field of life sciences, HiDream and its partner enterprises, through unified representation of multiscale biological imaging, achieve the understanding and prediction of cell evolution, providing a forward reference for drug response evaluation and screening out the more worthwhile solutions to validate before experiments.