0.06 seconds.

That's VLX-Flow's minimum latency for processing a single video stream. For humans, 0.06 seconds is roughly one-third of a blink. In that time, the model completes understanding a new frame, updating its environmental awareness, and judging whether a reaction is needed.

On July 1, Om AI released the VLX series -- the world's first edge streaming multimodal model for the physical world. It did something no previous vision multimodal model had done: instead of "watch a video clip then answer," it "understands while watching, and proactively acts when necessary."

From "Frame-by-Frame Screenshots" to "Streaming Perception"

How do existing video understanding models work? They slice video into frame-by-frame images, feed them into the model for offline processing, then output a result. That's like a person who opens their eyes every few seconds, looks at one photo each time, then closes their eyes to analyze it -- you can imagine how clumsy that perception would be, unable to cope with dynamically changing environments.

VLX's approach is fundamentally different. It proposes a "streaming multimodal" architecture: visual information enters the model as a continuous flow. Through incremental encoding and cached inference mechanisms, the model continuously observes the environment like a human, absorbing new frames as they arrive and responding instantly when queried. This doesn't correspond to a "better video Q&A experience" -- it's a qualitative leap in AI's ability to work autonomously.

The VLX series consists of three models: VLX-Flow handles continuous perception (incremental encoding + cached inference), VLX-Seek handles precise localization (turning coordinate generation into region retrieval -- not "guessing coordinates" but "selecting regions"), and VLX-Go handles action execution (converting visual understanding directly into robot-executable motion trajectories, not outputting text suggestions). The three work together to close the full loop of "continuous perception -> precise localization -> action decision" on edge devices.

Model sizes range from 0.6B to 10B parameters -- not cloud-scale models compressed onto edge devices, but redesigned from the architecture level for edge embodied AI.

Physical AI at Mid-2026

At CES earlier this year, Jensen Huang mentioned "physical AI" 17 times in his keynote, declaring that "physical AI's ChatGPT moment has arrived." Half a year later, signal density in this field is indeed exploding.

On the funding side: Nuwa Robotics, founded by former Baidu Autonomous Driving and Robotics Lab director Dr. Yang Ruigang, closed 50 million RMB across seed and angel rounds just 4 months after founding; Ganzhi Era, founded by Suzhou University Professor Liu Ruiyuan making multimodal electronic skin for robots, raised a 10-million-level angel round from Songhe Capital; Youai Zhihe released new embodied AI products targeting 10,000 industrial sites empowered in 3 years. Wayve (autonomous driving AI) initiated an $85 million employee share buyback at an $8.5 billion valuation. In the first half of the year, China's embodied AI sector saw 288 funding events totaling over 46 billion RMB.

On the technology side: AgiBot's 15,000th general embodied robot, the Genie G2, rolled off the production line; Unitree went from filing to IPO approval in a record 73 days on the STAR Market. As 36Kr reported, physical AI has made key progress across five dimensions: VLA models, world models, training grounds, hardware, and commercial closed loops.

But cool heads are needed. Physical AI's "ChatGPT moment" hasn't truly arrived yet. Robots perform increasingly well in controlled environments (factories, warehouses) but generalization in open environments remains limited. "Body"-level technologies like tactile sensing, force sensing, and motion control are progressing slower than the "brain" (VLA models). VLX solves the "eyes and brain" problem, but robots still need mature "hands" and "feet."

But the direction is clear. As AI moves from screens into the physical world, "streaming perception" is an architectural paradigm shift -- it lets AI start understanding the world through a continuous temporal dimension like biological organisms, rather than through discrete screenshots. This is the necessary path for physical AI to go from demo to product.

See you tomorrow.