Industrial Software Is Beginning to "Abandon Large Models": AI Is Shrinking and Becoming Everywhere

Many people assume that bigger AI is stronger AI.
But something very counterintuitive is happening in manufacturing –

industrial software is beginning to proactively "abandon large models."

In the last two days, a very typical signal has appeared.
The US embedded AI company embedUR announced that its growth core no longer comes from traditional software services, but from a platform called ModelNova: it provides a large number of small AI models that can be deployed directly on devices, specifically for logistics, medical devices, and industrial site control.

Note one keyword: deployed on devices.
Not in the cloud, not in data centres, but next to the machine.

Why?
Because factories and the internet have one essential difference –

the real world has no waiting time.

For a production line vision inspection, if the image is uploaded to the cloud and the result returned, even a one-second delay can cause an entire batch of products to be scrapped. So more and more manufacturing enterprises are moving AI from the cloud back to the site, letting models run directly on industrial PCs, robots, and even sensors.

The latest manufacturing quality research shows that AI quality inspection is achieving real-time decision-making through edge computing, correcting errors immediately during production rather than analysing them afterwards.
Behind this is actually a reversal of the industrial AI paradigm.

01.The Internet Pursues One Model Serving a Billion People; Manufacturing Needs Ten Thousand Models, Each Solving One Specific Problem

What the internet world pursues is:

one model serving a billion people.

But what manufacturing needs is:

ten thousand models, each solving one specific problem.

For example:
one model specifically judges solder joint defects;
one model predicts bearing wear;
one model optimises energy consumption;
one model understands equipment manuals.

The model does not need to be able to chat; it only needs to be stable, cheap, real-time, and reliable.

Even South Korea's latest industrial AI strategy has explicitly proposed embedding intelligence directly into the production site, rather than relying on centralised AI systems.

02.AI Begins to Be Used Like Industrial Components

So a new trend has emerged:

AI is beginning to be used like industrial components.

In the future, deploying AI may no longer be "launching a large model,"
but rather like installing PLC modules –

installing dozens of small models on a production line.

What does this mean?

What truly transforms manufacturing may not be GPT-level AI, but those "small model legions" that no one talks about.

Because what the industrial world pursues has never been intelligence, but determinism.

Large models change cognitive efficiency,
while small models are changing production efficiency.

So in the coming years, the biggest change in industrial software may not be AI getting bigger, but rather –

AI getting smaller, and then becoming everywhere.