The AI Bakery

Open Source, Open Weight and Closed AI Explained

Why is China giving away some of its best AI models? The answer becomes easier if we think about AI as a bakery.

The Closed Bakery

Imagine a bakery that sells an excellent cake. You can buy it and enjoy it, but you cannot enter the kitchen. You do not get the recipe, ingredients or baking process.

This is similar to closed AI. Users access the model through an app or API, while the company keeps control of the model weights and core technology. Leading proprietary AI services follow this general approach.

The Open-Weight Bakery

Now imagine another bakery. It gives you the finished cake and also gives you the recipe needed to make your own version. You can take it home, modify it and build something around it.

This is closer to open-weight AI. Models from Chinese AI labs such as DeepSeek, Qwen and Kimi have helped make this approach important globally. Developers can obtain released model weights, run models on their own infrastructure and adapt them, subject to each model's license.

Open Source Is Not the Same

The words “open source” and “open weight” are often mixed together.

A model may release its weights without releasing everything used to create it. The training data, complete training process or other components may remain private.

So AI openness is a spectrum. For companies, the license and actual usage rights matter more than the label.

Why Is China Doing This?

At first, the strategy looks strange. AI models cost huge amounts of money and computing power to develop. Why make powerful models widely available?

One reason is distribution. An open model can spread quickly through developers, universities, startups, cloud platforms and enterprises.

A second reason is ecosystem power. When thousands of developers build applications, agents, tools and specialized models around the same foundation, that model becomes more valuable.

A third reason is economics. Chinese AI companies do not have to compete only by spending more on computing. They can compete through efficiency, lower costs and wider adoption.

Open models also create external innovation. Developers around the world can optimize, fine-tune and build new applications around them. The original company benefits from a much larger ecosystem than it could create alone.

Open Does Not Mean Free Business

Giving access to model weights does not mean there is no business model.

Revenue can come from cloud computing, hosted inference, APIs, enterprise services, customization, support and commercial licensing.

The strategic idea may be simple: give developers greater access to the intelligence, then build businesses around the ecosystem.

Two Different AI Strategies

We may now be seeing two broad strategies.

The closed approach says: build powerful intelligence, control the model and sell access to it.

The open-weight approach says: distribute powerful intelligence, encourage others to build on it, and create a large ecosystem around it.

Neither approach automatically wins. They create different advantages and different risks.

And Then What Next?

The AI race may not ultimately be decided by which company has the highest benchmark score this month.

A bigger question is emerging:

Who will become the platform on which millions of developers, companies and AI agents build?

That may explain why China's open-weight strategy matters so much. What looks like giving AI away may actually be a long-term strategy for distribution, adoption and ecosystem influence.

A Simple Way to Remember It

Closed AI

You buy the cake and enjoy it, but the bakery keeps the recipe and controls the kitchen. You use the intelligence as a service, but you cannot download the core model or operate it independently. Leading proprietary AI services broadly follow this model.

Open-Weight AI

You get the finished cake and also the trained model weights. You can run the model in your own environment, adapt it, and build new services around it, subject to its license. But important details, such as complete training data or the full training process, may still remain private. DeepSeek, Qwen, and Kimi illustrate this broader open-weight approach.

Fully Open-Source AI

You get much more of the recipe, tools and process needed to study, modify and reproduce the system. The goal is broad transparency and reuse, although the exact rights still depend on the license and what the developer actually releases.

The easiest way to remember it:

Closed AI = use their bakery.

Open weight = build your own bakery.

Fully open = study and rebuild the bakery.