Why Is China Open Sourcing AI?

Chinese AI developers release open model weights to reach more users, encourage developers to build on their work and make deployment possible outside their own services. DeepSeek and Qwen provide examples. The commercial opportunity comes from adoption and the products built around a model, while users still pay for computing and integration.

When we recorded our Asiabits Podcast conversation with Tuo Liu, the founder of Robotuo, I asked about the people building products in Shenzhen. Our discussion moved from physical components to the software that makes them useful. Tuo described small teams using Chinese models because they could afford them and could choose where to run them. That is a useful starting point for understanding this part of China's AI industry.

I'm Thomas Derksen, and the question I keep coming back to in our China coverage is practical: what can a business do with the technology? A model release matters when someone can turn it into a product, a service or a better internal process. The download announcement is the beginning of that work.

What does open source mean for a Chinese AI model?

The first thing to establish is exactly what the developer has released and what its licence permits. Downloadable model weights let another organisation run the trained model. The training data, complete training process and every tool used to produce it may have different availability. Describing all of that with the single phrase open source can conceal differences that matter to a buyer.

Two established examples make the distinction easier to see. The DeepSeek-R1 repository states that its code and model weights use the MIT licence and support commercial use and modification. It also identifies underlying licences for distilled models derived from other developers' work. The Qwen3 repository says its open-weight models use Apache 2.0 and directs users to the licence files for the individual releases.

Those are specific model families with documented terms. They provide a better basis for a decision than assuming every product from the same country, company or brand has identical rights. A hosted chat application is another product again: the application can have its own account conditions and data handling.

For a company comparing suppliers, the useful record is the model's exact name, version, licence and intended deployment. Keep that record alongside the configuration used in testing. Otherwise an evaluation can drift as people change a model, a provider or a setting without noticing that the underlying proposal has changed.

Why would a company release something valuable?

Releasing a usable model can increase the number of developers and businesses building around it. That gives the developer distribution beyond the customers it could reach through a single paid application. This is a commercial interpretation of the strategy; it does not establish the internal motives or economics of every Chinese AI laboratory.

A small software company can try a model inside a product without first negotiating a bespoke relationship with the laboratory. Developers can share examples and build tools that make the next team's work easier. A company that has already integrated a model may later buy computing, support or other services connected to its use. The open release and a paid business can therefore coexist.

Tuo's account of Shenzhen adds a useful view from the builder's side. In our published conversation about Robotuo, he described affordability and the option to deploy locally as reasons smaller teams choose open models. He also discussed how accessible software helps teams in markets where budgets for paid AI services are limited. That is his experience of the founders he meets, rather than a measured global market share.

In an earlier LinkedIn post about the AI race, I made a related point: the model a business can deploy successfully deserves attention alongside the model that leads a public demonstration. Here the relevant comparison is the complete application and its economics. A downloadable model that requires an expensive operating setup can be a poor fit for a small company even when the licence itself costs nothing.

Does an open model make business AI free?

An open model changes the purchase and deployment options; it leaves the cost of running a useful application to be worked out. Computing, engineering time and ongoing operations still need a budget. The appropriate comparison is the cost of completing the business task to the required standard.

Consider an internal assistant that reads equipment manuals. A working service needs access to the correct documents, a way to retrieve relevant passages and a clear response when the documents do not contain an answer. Someone must decide which employees can see which manuals and what happens when a new revision arrives. The language model supplies only part of that service.

Testing should include difficult questions from actual work. Ask for an answer based on an obsolete document, ask a question with missing information and include terminology the business uses internally. Record whether the system finds the right evidence, whether its answer is useful and how much human correction it needs. A fluent response can still be operationally wrong.

Running a model yourself also requires someone to own updates and failures. For a team with suitable infrastructure and steady usage, that may be a sensible arrangement. A team with little engineering capacity may prefer a managed service. The decision becomes clearer when both proposals include the same workload, service expectations and maintenance responsibilities.

This is why a claim about a model being cheaper needs a defined comparison. A price per request, a licence price and a fully staffed internal service describe different costs. Compare them after writing down the task and the expected amount of work, rather than treating the smallest advertised number as the total bill.

How does this connect to robots and hardware in China?

Open models give hardware developers another software option, but a physical product still needs engineering beyond a language model. A robot must perceive its surroundings, control movement and handle the consequences of an incorrect action. The responsibilities differ substantially from generating a paragraph on a screen.

In our database of 7,000+ robotics companies, the snapshot checked for this article contains 7,320 company records. That is the size of the directory snapshot checked for this article, not a count of companies deploying Chinese open models. The records cover different parts of robotics. They are a reminder to look beyond the company supplying a complete machine when researching who can deliver a working system.

In Tuo's account, global founders come to Shenzhen for access to hardware production and for the people who can help build a product. He described software developers contributing to products made by hardware teams. Shared models can make that collaboration easier because teams can work with a common starting point. The physical design, available components and the target market still shape the final product.

For a buyer, the useful question is which task a proposed AI component performs. It might help with a spoken interface or interpreting a request. That does not establish how safely the machine moves or how reliably it completes a physical task. The supplier needs to explain the boundary between the conversational system and the controls responsible for physical operation.

Our guide to embodied AI and China's robot builders develops that distinction. Our account of founders building in Shenzhen provides the wider context from the Robotuo conversation. Together they help turn a broad AI question into a more specific product discussion.

Can a company keep its data under its own control?

Local deployment can give a business more control over where its model runs and where its inputs are processed. It does not by itself settle every question about the application. The complete data flow matters, including any external tools the application calls and any information retained in logs.

A proposed internal assistant may still use an external search service, a hosted document store or a monitoring service. Each connection deserves an explanation. A buyer should be able to trace a sample document through the system and identify who can access the input, the output and the stored history. That exercise is more informative than a general assurance that the model can run locally.

Start an evaluation with material that is appropriate for testing. Define the allowed users and the actions the application may take. If it can send a message, change a record or issue an instruction to another system, decide where a person must approve the action. Those boundaries should be visible in the demonstration and in the proposed operating process.

A model's country of origin alone cannot answer these deployment questions. The buyer needs the actual provider, hosting arrangement and application design. Equally, the ability to download weights does not establish that a particular supplier has built a dependable service around them. Both the software choice and the operating arrangement need evidence.

How should a business evaluate the opportunity?

Choose a task with an observable result and compare complete proposals against that task. An internal document assistant, a product support tool and a robot's spoken interface have different requirements. Starting with a single use case makes it easier to see whether an open model changes the business case.

Write down a small set of representative examples before asking suppliers for demonstrations. Keep the inputs and scoring consistent between proposals. Include the cases where the correct response is to request clarification or refer a matter to a person. A good evaluation shows where the application works and how it behaves when it reaches its limits.

Then ask who will maintain it. The proposal needs an owner for the model configuration, the application and the source material. Establish what evidence will be available when an answer is disputed. Include the cost of that support in the comparison. A promising demonstration becomes easier to assess when the operating responsibilities are explicit.

Chinese open models are useful candidates for this process because they expand the available choices. Their suitability depends on the actual release, the task and the team operating it. We cannot infer a laboratory's long-term commercial success from a licence or infer a customer's savings from a download. Those questions require evidence from the application over time.

Asiabits connects international companies with China's technology and robotics industry through research and visits. For teams assessing an AI-enabled physical product, our bespoke China visits help organise discussions around the application, the relevant builders and what a demonstration can establish. Bring a defined task, and the conversation with a supplier becomes far more useful.

Frequently asked questions

Why do Chinese AI companies release open models?

An open release can help a model reach more developers and become part of other companies' products. This is a commercial interpretation of the strategy; individual developers have different business models.

Are DeepSeek and Qwen available for commercial use?

The DeepSeek-R1 repository describes MIT-licensed code and weights with commercial use. Qwen3 describes Apache 2.0 open-weight releases. Check the exact release and any underlying model licences for the intended application.

Does running an open model eliminate AI costs?

Computing, application development and maintenance still have costs. Compare the complete cost of doing the same task to the required standard.

Can open models run without sending data to a public chatbot?

A local deployment can process model inputs within infrastructure controlled by the operator. Check the rest of the application's connections and logging to understand the complete data flow.

Can an open language model control a humanoid robot?

A language model can be one component of a robot application, for example in its spoken interface. Physical operation also requires perception, control and a defined safety design; a conversational demonstration does not establish reliable task performance.