NAIROBI, Sept. 1 (Xinhua) -- Chinese artificial intelligence (AI) models are gaining ground among Kenyan developers and technology companies, as their lower costs, competitive performance and open-source nature offer an attractive option for a market seeking affordable and customizable AI solutions.

A recent analysis by The New York Times showed that Chinese AI models accounted for roughly half of AI use on OpenRouter, a platform offering about 400 AI models, up from less than 25 percent a year earlier.

On Hugging Face, another major AI model platform, Chinese AI models accounted for 19 of the 25 most-downloaded open-source systems.

In Kenya, developers and technology companies are increasingly turning to Chinese AI models such as DeepSeek and Qwen for applications ranging from legal technology and business intelligence to document processing and coding.

Michael Michie, founder and chief executive officer of EverseTech, an AI startup focused on local GPU infrastructure and enterprise AI applications, said Chinese models have become the most widely used models in many of the company's projects.

"Most of my clients choose the Chinese one," said Michie, who was recently appointed the responsible AI governor for Kenya, citing cost and computing efficiency as major factors.

Kenya has a limited supply of GPUs, making models that require fewer computing resources easier to deploy on local infrastructure, he said.

"They perform just as well, sometimes even better," Michie said. "In the past, Chinese models were maybe six months or a year behind in terms of capabilities, but now they do the same complex tasks that frontier models would do."

He noted that Chinese models also consume fewer tokens for some tasks, further reducing the computing resources required.

Michie's view is echoed by Mbugua Njihia, a Kenyan technology expert, who said the cost advantage matters especially in Africa, where AI developers face a much more complex linguistic environment than in many other markets.

Against this backdrop, the availability of open-weight models gives African developers greater scope to adapt AI to local needs, Njihia added.

By downloading and retraining the models with relatively small datasets, developers can customize them for local languages, specific industries and other Africa-specific applications without building a model from scratch, Njihia said.

The practical value of such flexibility can be seen in JibuDocs, a document-processing tool developed by entrepreneurs Sentai Simons and Daniele Pisani using a Chinese AI model.

With a Chinese AI model, the tool can process large volumes of documents, extract and organize information, and generate summaries, helping users find specific information without having to manually read through every document, according to Pisani.

For Simons, one of the most important advantages of open-weight models is the ability to use a company's own data to further train and fine-tune a model for specific tasks.

Simons, whose team uses both DeepSeek and Qwen, said Chinese AI models "are very good and can be a very good starting point," he said, adding that, more importantly, it allows users to customize it and continue training it.

That combination of performance and openness is also attracting Kenyan technology company Qhala, which began experimenting with DeepSeek in the second half of 2024 and put more production systems into operation in early 2025 following the release of DeepSeek R1.

Shikoh Gitau, Qhala's chief executive, said DeepSeek's technical research has also influenced how its team approaches AI research and development.

Its research papers on areas such as mixture-of-experts architecture and reinforcement learning have helped the team understand advanced model development and rethink assumptions about the resources needed to build capable AI systems, according to Gitau.

The company is also looking at ways to use DeepSeek to help develop cost-efficient models optimized for African languages and content, Gitau added.

Kenya was one of the 29 founding members of the World Artificial Intelligence Cooperation Organization, which was established in July at the World Artificial Intelligence Conference in Shanghai.

Michie said he hoped Kenya could strengthen partnerships with China to expand its computing capacity and establish exchanges that would enable Kenyan developers to learn from Chinese experts in building world-class AI models.

Njihia said Chinese AI models could also be useful in sectors where Africa faces shortages of specialized expertise and infrastructure.

Healthcare triage and diagnostic-support tools trained on local disease patterns and language data could help extend limited clinical expertise to rural communities, provided that data governance remains local, he said.

Disaster prevention is another area where such technology could have practical value, with flood and drought modeling combined with early-warning systems in local languages, Njihia added.