China: 41% of downloads and a question of trust

Bajío Ventures · October 2026 · Sovereign AI, 4 of 5

In January 2025 a laboratory in Hangzhou, DeepSeek, released its R1 model under the MIT licence. That was the start of the wave. Less than two years later, these are the numbers:

• Chinese models accounted for 41% of downloads on Hugging Face in the year before the platform’s March 2026 report, the largest share of any country.
• Alibaba has more derivative models than Google and Meta combined; its Qwen family alone has more than 113,000.
• The Tony Blair Institute reports, with data from The ATOM Project, that Chinese models account for 50% to 60% of new open-model adoption, against 25% to 30% for the United States.
• As of 5 October 2026, the five top-rated open models in the Artificial Analysis index are Chinese.

The models

Model	Laboratory	Release	Parameters (total / active)	Licence
MiMo-V2.6-Pro	Xiaomi	21 Sep 2026	1.02 trillion / 42 billion	MIT
GLM-5.3	Z.ai (Zhipu)	28 Aug 2026	753 billion	Custom, with security review
Kimi K3	Moonshot AI	27 Jul 2026	2.8 trillion / 104 billion	Custom (Kimi K3 License)
DeepSeek V4.1 Flash	DeepSeek	10 Sep 2026	552 billion / 8 to 16 billion	MIT
Qwen3.8	Alibaba	14 Aug 2026	2.4 trillion / 95 billion, and one of 27 billion	Custom for the large model; Apache 2.0 for the 27 billion model
Hy4 preview	Tencent	Aug 2026	770 billion / 49 billion	Apache 2.0

Three things the table shows

1. These are data center models. The Kimi K3 weights take up about 1.4 terabytes after 4-bit quantization. The vLLM guide asks for at least 8 latest-generation GPUs and several nodes for production traffic; Moonshot recommends 64 accelerators or more. Nobody runs that in a plant. The Chinese model that does fit in an industrial server is the 27-billion-parameter Qwen3.8, under the Apache 2.0 licence.

2. Licences tightened in 2026. DeepSeek keeps the MIT licence, and Z.ai used it up to GLM-5.2. This year three laboratories put conditions on their flagship model:

• Large Qwen3.8. Companies that offer models as a service or AI work assistants and exceed USD 50 million in revenue over 12 months need a separate licence. Products with more than 100 million monthly users or USD 20 million in monthly revenue must display the model name.
• Kimi K3. Model-as-a-service providers with more than USD 20 million in annual revenue need a separate agreement with Moonshot. The same rule on displaying the model name applies as for Qwen.
• GLM-5.3. Companies with more than USD 10 billion in revenue over 12 months that want to host the model must pass a Z.ai security review before any commercial use.

The thresholds target clouds and resellers, not a plant that uses the model for itself. But the trend matters. The New Stack puts it this way: if the next version of GLM ships the same way, open weights will start to look less like a gift and more like a distribution channel with conditions.

3. Opening up is their way of selling. An analysis in Forbes describes the strategy: release for free a capability that sold at frontier prices a year ago, in markets where Chinese companies face trust and procurement barriers to selling closed services to Western companies.

The question of trust

Three risks are often mixed together. Separate them.

Risk	Does it apply to an open model on your own server?
The data travels to China	No. The model runs locally and needs no connection. It does apply if you use the laboratory’s API.
The model carries biases or content filters from its origin	It can. What the model learned is in the weights. It matters little for reading maintenance orders and a lot for political or legal analysis.
The customer or the regulator objects	It depends on the customer. This is the real risk for an exporter.

The third risk is not hypothetical for Mexico. According to September 2026 reports that cite the Wall Street Journal, AI hardware overtook automobiles this year as Mexico’s main export to the United States: USD 83 billion in servers in the first half, 94% of it bound for that country. Washington is seeking rules of origin with higher North American content for that hardware, similar to the 75% that applies to automobiles.

In that context, a Bajío supplier that serves US automakers or aerospace customers must ask before deploying. It is the same conversation that already exists about Chinese-origin electrical equipment in a plant’s electrical room.

What it means for a Bajío plant

• For internal use, the mid-sized Chinese model is a valid technical option. The 27-billion-parameter Qwen3.8 has the same licence as Gemma 4 or gpt-oss.
• Run the model, not the API. The sovereignty advantage only exists if the weights are on your equipment.
• Document the decision. Record which model, which version and where it runs. It is the first question in a customer audit.
• Keep a second model tested. If the customer objects, the switch must take days, not months.

In this series: Guide: what sovereign AI is · 1. Aleph Alpha · 2. Mistral · 3. United States · 4. China · 5. Spanish-language models and Mexico

Sources: Hugging Face, State of Open Source, Spring 2026 · Tony Blair Institute, Open Source, Feb 2026 · Artificial Analysis, open models · Wikipedia, Xiaomi MiMo · The New Stack, GLM-5.3 licence, Aug 2026 · Devoriales, Kimi K3 · vLLM, Kimi K3 guide · Hugging Face, DeepSeek R1 · DataNorth, DeepSeek V4.1 Flash, Sep 2026 · SQ Magazine, Qwen3.8 licences, Aug 2026 · Forbes, Tencent Hy4, Aug 2026 · Seoul Economic Daily, rules of origin for AI hardware, Sep 2026

Bajío Ventures — a Mexican energy-infrastructure investment group based in Guanajuato: capital, a data platform on the region’s industrial base, and a German industrial channel, with its own advisory presence in Berlin.