The “world model” is becoming 2026’s most abused technology label. Video generation models claim the title. Vision-language-action models claim it. JEPA claims it too. Each has found a comfortable spot under the term’s wide brim.
Which raises the question that nobody in the label rush wants to answer: if everything is a world model, what isn’t?
Li Yiming — an assistant professor at Tsinghua University’s Institute for AI, born in 1997, already advising PhD students — delivered three unusually sharp verdicts in a recent deep-dive interview. First: VLA models are not native world models, because their representation lives in the discrete space of language, not the real world — and language is a heavily compressed abstraction of world rules, built for communication between humans, not for machines to understand physics. Second: video generation models are not native world models, because the reasoning process is not native — their pixels fit the surface appearance of the world but cannot guarantee the geometric and physical consistency that policy learning for complex tasks requires. Sora can generate a beautiful rowing video; it cannot train a robot to row. Third: JEPA is not a native world model either, because it predicts states and cannot output actions — perception without action is not a complete loop with the physical world.
Those three sentences peel the bubble off the 2026 world-model land grab with surgical precision. And the market has voted: Li’s startup, Liqing Intelligence, closed a seed round in the hundreds of millions of RMB within two months of founding — with Shunwei Capital, Sequoia China, Hillhouse Ventures, and Fengrui Capital joined by strategic investors including Zhiyuan Robotics and Lingxin Dexterous Hands. A nine-figure endorsement of exactly these judgments.
The Definition Problem Is Actually a Systems Problem
Li’s persuasiveness rests on one cognitive premise: a world model is not a model. It is a system. He illustrates with the historical novel The Litchi Road: to transport fresh lychees from Lingnan to Chang’an, the minor official Li Shande had to solve preservation, relay stations, routing, and supplies as one interlocking system. No complete system, no delivery — a fast horse alone transports nothing. A world model is that horse: one technical route among several, worthless without the data collection, hardware deployment, and physics engines around it.
The analogy lands on the industry’s biggest blind spot. Most companies treat “train a model that generates video or 3D” as the goal. Li’s stated goal is a Physical AI infrastructure driven by data and physics in dual wheels — an end-to-end system spanning perception, reasoning, decision, and action output, with the model as one component inside it. His phrasing is pointed: “We don’t want the ‘world model’ label. What we’re doing is solving Physical AI.”
Why VLA Isn’t ‘Worldly’ Enough: Three Dimensions
Li’s critique of the VLA route deserves careful reading from anyone following embodied AI.
Language is not a modality; it is a communication interface. VLA’s core flaw is treating language as the central representation. Language is a second-order compression of the world — it discards physical detail (the center of mass of an apple, its softness, its friction coefficient) and carries human cognitive biases baked into every language. Li’s division of labor: language models handle machine-to-human interaction; world models handle machine-to-world interaction. Different functions, no substitution possible.
Tokenization is the real bottleneck. For world models the central challenge is not parameter count or compute — it is how to efficiently tokenize the physical world. How visual, tactile, and force observations compress into token sequences a model can reason over sets the ceiling on everything else. Li’s team has already surpassed Meta’s DINOv3 on visual tokenization, and that is only the start. The governing principle: language tokens are discrete and conventional; physical tokens must be continuous and physically consistent. That structural difference is what separates world models from language models.
The data-efficiency leap. The team runs a “1% real-robot data principle”: train policy models with 1% of the real-machine data others use, and match their success rates. The method is simple and clever — collect a small batch of real-robot data, align the robot’s real state transitions with the world model’s simulated state transitions, and backpropagate that loss to refine the world model. After alignment, most of the robot’s practice happens inside the world model. Learning to cut an apple used to destroy hundreds of apples; now it destroys ten, with the remaining repetitions virtual. That is not a cost saving. It is the precondition for Physical AI to cross from demo to deployment at scale.
The ‘Litchi System’: A Three-Component Physical AI Stack
Liqing’s technical route decomposes into three components, which Li names the litchi system — after the novel, because fresh lychees reach the capital via systems, not fast horses.
Component one: the data pipeline. Self-developed capture devices — including full-palm tactile gloves — push per-unit costs from dollar scale down to RMB scale, and data volume from the industry-typical hundreds of thousands of hours toward millions and tens of millions. One counterintuitive bet: human data scales more easily than robot data. China has hundreds of millions of people working on production lines and living in households; people wearing capture devices generate data far faster than robots being teleoperated.
Component two: a native world model. Not VLA, not video generation, not JEPA — a single model with perception, reasoning, decision, and action integrated end to end. Pretraining runs self-supervised on predicting the next state and the next action; post-training packages the model as an interactive simulation environment.
Component three: a differentiable physics engine. The most underestimated capability in the stack. The engine must model soft bodies and fluid dynamics efficiently enough for robots to run reinforcement learning inside it — and it must be differentiable, so loss can flow backward and the world model can keep improving from what the engine reveals.
The three components close a loop: the data pipeline produces training data; the world model learns physical laws; the physics engine supplies the post-training environment; better models produce better robot behavior; better behavior generates higher-quality data. The endgame is to package the whole system as general Physical AI infrastructure — what iOS is to mobile apps — so physical-operation tasks can be developed and deployed at scale.
The 2028 Milestone — and Who Gets Eliminated First
Li projects 2028 as the milestone year when Physical AI lands at scale, resting on two trends: data collection volume jumps within twenty-four months as China’s industrial workforce carries capture devices (millions of hours of training data), and motor density plus hardware costs keep improving fast. More useful than the date, the framework predicts which company shapes get eliminated along the way. Model-only companies — no data pipeline, no physics engine, just a model trained to imitate a world model — are the fast horse without relay stations; they deliver nothing. Hardware-only companies — bodies and control solved, but no model and no engine — produce robots that execute preset programs in real environments with no understanding. Data-collection-only companies — mountains of raw data with no ability to turn it into model training or simulation — own smart asphalt that no car drives on.
The only company shape that runs the full loop, Li argues, is “World Model as Service”: not a robot-body company, not a model company, but a company that delivers integrated software-hardware systems to customers — automatically matching the optimal hardware configuration to each scenario and budget, working out of the box.
What To Do With This
If you invest in Physical AI or embodied intelligence, ask every target three questions: Is your tokenizer trained in-house or borrowed? Is your physics engine differentiable? What is your real-robot versus synthetic data ratio? Companies that cannot answer are probably demo shops wearing world-model costumes.
If you are a founder, look past parameters and benchmarks. Data-pipeline and physics-engine capability may matter more than training a bigger model — in this stack, full-stack coverage is the moat.
If you are a technical decision-maker, start thinking about physical tokenization for your own domain. What you will eventually need is not an API to call a world model but an understanding of how the physical world gets represented and reasoned over. The gap between those two mindsets decides whether your product closes its loop before 2028.
Liqing’s nine-figure seed round is not just a bet on a Tsinghua-affiliated startup. It is a value judgment on the question “what is a real world model” — and a warning that in this race, the horse was never the point.
Source: 36Kr, “Intelligent Emergence” exclusive report.
