Chinese Open Models and the One-in-Five Claim

Not as stated: there is no reliable, clearly identified evidence that Chinese open models have won over one in five of “the informations subscribers.” The phrase appears to contain a typo or an unclear reference, so the claim cannot be verified without knowing which subscribers, survey, product, or time period it means.

What the claim would need to establish

A statement that a model has won “one in five” subscribers is a market-share claim. To assess it, readers would need to know the denominator (all subscribers to what service?), the measurement period, the meaning of “won” (paid subscriptions, sign-ups, usage, or preference), and which models count as Chinese and open. Without those details, a percentage sounds precise but cannot be checked.

“Informations subscribers” is not a standard market category in the AI industry. It could be a misspelling of “information subscribers,” a reference to subscribers of a particular publication, or an incomplete phrase. It should not be silently interpreted as a named company, survey, or user group. If the intended reference is to subscribers of a specific information service, that service’s own figures or a well-described independent study would be needed.

What “Chinese open models” means

The phrase usually refers to language or multimodal models developed by organizations based in China whose weights or other components are made available for outside use. The label “open” is used inconsistently. Some releases provide downloadable model weights, while code, training data, or detailed training methods may remain unavailable. A model can therefore be open in one practical sense without being fully open in every sense.

Examples of Chinese-developed model families that have received international attention include Qwen, DeepSeek, and models from Zhipu AI’s GLM family. Their releases differ in model size, capabilities, access terms, and openness. Naming a family does not establish how many subscribers use it, and the existence of a downloadable model is not evidence of paid adoption.

Why subscriber counts are difficult to compare

AI use is split across several channels. People may access a model through a developer’s own chatbot, a cloud provider, a third-party application, or a self-hosted installation. A subscriber to an application might never know which model answered a request; conversely, someone running a model locally may not be counted as a subscriber anywhere.

Companies also report different measures. “Users” may mean registered accounts, monthly active users, paying customers, or API customers. These measures are not interchangeable. Downloads and model-hub activity can indicate interest, but they do not reveal how many people use a model regularly or pay for access. API request volumes can show deployment activity while leaving the identity and number of end users unknown.

• Direct subscriptions: paid accounts for a provider’s own product, if publicly reported.
• Third-party use: access through another company’s service, where the underlying model may not be disclosed.
• Self-hosted use: downloaded models running on a user’s own hardware or infrastructure.
• Popularity indicators: downloads, rankings, or benchmark results, which measure something other than subscriber share.

What adoption evidence can and cannot show

Public releases, technical documentation, model downloads, developer discussions, and benchmark evaluations can show that a model is available and attracting attention. They can also help compare performance on particular tasks. But none of these is, by itself, a count of subscribers. Benchmarks measure performance on selected tests, not market penetration or customer satisfaction.

To substantiate a one-in-five figure, a report should identify its source and explain the sample. For a survey, useful details include the number and location of respondents, how they were recruited, the exact question, and whether respondents could choose multiple models. For a company’s subscriber statistics, the reporting period and definition of “subscriber” matter. For market share, the report must define the market and explain how it handles bundled services and anonymous model use.

A percentage may also refer to preference rather than actual subscriptions. If one in five survey respondents says they prefer a Chinese model, that does not mean one in five subscribers has switched to it. Preference, trial, active use, and payment are separate outcomes.

Why developers consider open-weight models

Open-weight models can give developers more control over deployment. Depending on the model’s license and technical requirements, an organization may run it on its own infrastructure, adapt it for a task, or reduce dependence on a single hosted service. This flexibility can be valuable for experimentation and for applications that need customized workflows.

Those advantages come with trade-offs. Running a model requires suitable hardware or cloud infrastructure, operational expertise, and ongoing maintenance. Performance can vary by task and language, and hosted services may offer features—such as managed scaling or integrated tools—that a downloaded model does not provide by itself. License terms also vary, so users need to check the conditions for their intended use rather than assume every model permits every deployment.

How to evaluate a market-share statement

Before repeating the claim, look for the original source and ask:

• Who exactly are “the subscribers,” and what service are they subscribed to?
• Does “one in five” describe paid customers, active users, survey preferences, or something else?
• Which models are included, and what definition of “open” is used?
• When was the data collected, and does the source publish its method?
• Can the figure be independently checked, or is it an unqualified promotional statement?

If these questions have no clear answers, the responsible conclusion is that the claim is unverified—not that the number is necessarily false. A properly sourced figure could be meaningful for a defined service or survey, but it should not be generalized to all AI subscribers.

Frequently Asked Questions

Do Chinese open models have one-fifth of AI subscribers?
No general, verifiable figure establishing that share is identifiable from the wording of the claim. It needs a named subscriber group, a source, and a clear definition of the measure.

Which Chinese open models are commonly discussed?
Qwen, DeepSeek, and the GLM family are examples of Chinese-developed model families that have drawn attention. Their availability, terms, and degree of openness differ by release.

Does downloading a model count as subscribing to it?
Not necessarily. A download may be free and used locally, while a subscription generally refers to paid or account-based access. Reports should state which measure they use.

Are open-weight models fully open source?
Not always. Weights may be available while training data, code, or other components are not. The license and release details determine what users can inspect, modify, and distribute.

What evidence would prove the one-in-five claim?
A credible source would define the subscriber population and time period, explain how the figure was measured, identify the relevant models, and distinguish actual subscriptions from usage, downloads, or stated preference.

Related

• Google Sheets Combine Date and Time: How It Works
• Apple Intelligence: Is It Worth Buying a New iPhone?
• Exchange Administrator Made a Change? What It Means