The physical-AI track in the summer of 2026 has grown dizzyingly crowded.
VLA (Vision-Language-Action), world models, JEPA, video-generation simulation — every route has its advocates, and every faction claims to hold the key to physical intelligence. The industry debates whether “VLA is dead, world models have arrived” or “the three will eventually merge,” while a more fundamental question goes unasked: what architecture does AI in the physical world actually need?
Om AI’s recently released VLX series offers a rare sample of an answer. This is not another cloud model trimmed down for edge deployment — it is a native architecture designed for the physical world from day one, called streaming multimodality. For the five years the industry chased generative dialogue, almost nobody understood what Om AI was doing. Until now.
The Architectural Paradox: Text Thinking for a Video World
Start with a counterintuitive fact. In 2023, Om AI connected a multimodal model trained on general media data to a bank of surveillance cameras. The team was pessimistic — the surveillance field already had many mature specialized small models. The result upended expectations: the general model, which had never seen surveillance footage, outperformed the veteran small models trained on surveillance data for years.
The episode exposed a deep problem: last generation’s “small model plus specialized data” paradigm in computer vision fundamentally lacked generalization. Change the scene, the lighting, the camera angle, and the model failed. General multimodal models, having seen enough of the world’s variety, showed stronger adaptability in open scenarios.
But general multimodal models carry their own flaw — they were designed for the digital world. The mainstream approach today is “offline frame extraction”: video gets sliced frame by frame and processed as images; inference is batch, discrete, question-and-answer. Fine for cloud chat scenarios — but the physical world demands continuous perception. The camera is always watching, the drone always flying, the robot always moving. You cannot wait for a user to ask “what just happened” before processing.
This is the architectural paradox Om AI founder Zhao Tiancheng saw back in 2019: using text-mode thinking to understand a video world is fundamentally wrong. A language model’s core computation unit is the token, processed linearly as strings. But physical-world input is a video stream — not a string of equal-length characters but a continuously flowing, highly redundant signal. Most frames contain nothing new, yet the critical event can appear in any one of them. Approximating a video stream with text tokens is like transcribing a live broadcast on a typewriter — slow and dim.
VLX’s Choice: Flow, Seek, Go — Three Capability Layers, Not Three Models
Om AI’s answer is a three-layer architecture with plain names: VLX-Flow, VLX-Seek, VLX-Go. Flow is the continuous perception layer. Video flows into the model like water; the AI watches constantly without waiting for a question. It sounds simple but is technically brutal — the model must process video streams on-device at extremely low latency while automatically identifying which frames deserve attention and which to skip. That is not “compress a big model until it fits on the edge” — it demands redesigning, at the architecture level, how a model processes time-series information.
Seek is the precise localization layer. On top of Flow’s continuous perception, Seek answers “where is it” — geometric-spatial understanding. Not “there is a cat in the frame,” but the cat’s exact position and pose in three-dimensional space. The VLX-Seek 1.5 release at WAIC specifically upgraded the architecture for embodied scenarios — drones, robots, robot dogs — with systematic gains in accuracy and inference efficiency. Go is the action-decision layer. Knowing “what” and “where,” Go decides “how”: which way the robot walks, at what angle the arm grips, how the drone adjusts its flight path.
The critical distinction: these three layers are not three independent models stitched together — they are three inseparable capability layers on the same video stream. Flow continuously outputs semantic understanding, Seek localizes on that understanding, Go decides on that localization. They share one underlying architecture and one data stream, unlike traditional pipelines that split perception, localization, and decision into three disconnected stages. The design carries an underrated advantage: the data flywheel. When cameras, drones, and robots all run the same model architecture, every terminal’s data feeds all the others. A robot dog has little training data of its own — but it inherits experience from millions of cameras and PCs. That is Om AI’s “one brain, many bodies.”
“Edge-Native” Is Not Optimization — It Is an Architectural Choice
The most important concept for understanding VLX: it is not a shrunken cloud model; it is a native edge design. That sounds like semantics, but the difference is enormous. Cloud models assume unlimited compute, tolerable latency, and stable networks. The usual edge-deployment path: train a large model, then compress it to edge-runnable size via distillation, quantization, and pruning — “compression thinking,” subtracting from a perfect-but-expensive solution. Om AI ran it in reverse: the architecture was designed from day one around edge constraints. How small must the model be to run on a camera chip? How fast must inference be for real-time drone control? How memory-frugal to fit a wearable? These constraints are not optimization targets — they are the starting points of the architecture. The result: the VLX series achieves near or better-than-cloud performance in edge scenarios at parameter scales far below cloud models. Not a victory of engineering optimization — a victory of architectural choice. Once you stop treating the cloud as the default starting point, many “impossibilities” turn out to be merely “path dependencies.”
Zhao Tiancheng noted a key insight in interviews: physical-AI terminal scenarios are actually China’s advantage. “In purely digital spaces, everyone’s products are strong. But in physical terminal scenarios, China has the edge.” The logic: China has the world’s most complete hardware supply chain and the largest terminal deployment base — and as AI descends from cloud to ground, that infrastructure becomes a data advantage.
From Passive Response to Active Adaptation: the Physical-AI Paradigm Shift
The most valuable part of Om AI’s story is not the technical detail — it is the paradigm shift it reveals. For a decade, AI’s core capability in the digital world was passive response: user asks, model answers; user commands, model executes. That is the substrate of chat and generative AI. The physical world demands active adaptation — the camera decides for itself when to alert, the drone judges its own flight path, the robot senses environmental change and adjusts its motion. The difference is not a different tech stack; it is a fundamental change in AI’s relationship with the physical world. For passive-response AI, the world is static — no time pressure between input and output. For active-adaptation AI, the world flows — every millisecond of delay can have physical consequences. A wrong character in code is trivial; a robot freezing mid-motion and falling is serious. A self-driving car failing while moving is unthinkable.
This explains Om AI’s insistence on distributed intelligence. “Having everything controlled by one brain is honestly terrifying — like a sci-fi villain: seize the one AI and every terminal becomes a tool for hijacking humanity,” Zhao said. He wants every physical terminal to carry independent, local, fast-responding intelligence — like humans, where each brain is autonomous. The perspective is notable: while the industry debates “cloud model versus edge model, which is stronger,” Om AI’s answer is that the endgame is not either/or but distribution — cloud for global reasoning and long-term planning, edge for real-time perception and rapid response. The two are not substitutes; they are a division of labor.
Physical AI’s Cambrian: Who Is Accumulating Real Structural Advantage
Om AI has reached revenue measured in hundreds of millions of yuan — rare among physical-AI companies. Most are still burning R&D money without product-market fit, while Om AI has closed commercial loops on mature hardware: cameras, PCs, industrial IoT. Behind it sits a structural advantage worth noting: the time gap. Zhao began multimodal research in 2018 and multimodal model training in 2021. When the industry suddenly recognized multimodality’s value in 2024, Om AI already held three to four years of technical reserves, scenario exploration, and market education. “Our research started much earlier — three or four years ahead of others.” In fast-iterating fields, three to four years is a technology generation. And before technical routes converge, early accumulation is not just technology — it is the data flywheel and scenario understanding: knowing which scenarios genuinely work, which customers pay, which problems look small in the lab but are fatal in production.
Zhao compares physical AI today to deep learning in 2016: “Back then deep learning had just broken through, and nobody knew what the single convergent end state would be.” Physical AI is at the same stage — technical routes blooming, the zero-to-one breakthrough made, the one-to-hundred scale-out not yet begun. In this phase, the scarcest capability is not “picking the ultimate route” but getting the data flywheel spinning first — because whatever the end state looks like, data accumulation and scenario validation cannot be skipped. Om AI has millions of cameras, drones, and robots already running, feeding real business data back into model iteration. Once that flywheel turns, what challengers must catch up to is not just model capability but data scale and scenario depth.
Closing
Om AI’s VLX series is not physical AI’s final answer. Zhao himself concedes the field is “a building with only the frame up — many bricks still missing,” and no single model solves everything. But its value lies in providing an architecture-choice sample: while the mainstream narrative says “solve everything with a bigger model,” someone chose the harder road — designing from the physical world’s nature upward rather than trimming down a cloud model. The endgame is far off. But one thing grows clearer: in the physical world, architecture choice matters more than model scale — because physical constraints of latency, power, safety, and reliability are not solved with more compute; they demand different architectures to accommodate. And the companies that understood those constraints from day one are accumulating a structural advantage that is very hard to chase down.
