Why is China ahead in open-source AI?

In March 2026, Hugging Face reported that Chinese models took 41% of downloads on its platform over the past year, the largest share of any country and more than the US. The turn came after DeepSeek released its R1 model in January 2025 under the MIT licence. Since then, Chinese companies have published far more open models and updated them often, and developers build on them: most trending new models of 2025 were Chinese or derived from a Chinese model.

What does the Hugging Face data show?

The report State of Open Source on Hugging Face: Spring 2026, by four Hugging Face researchers, looks at activity on a platform that hosts more than 2 million public models. On geography it finds:

• In the past year, Chinese models took 41% of downloads, the largest share of any country. China overtook the US in both monthly and overall downloads.
• Before that, the US and Western Europe led through large industry labs: Google, Meta, OpenAI and Stability AI.
• Of the models created in 2025, most of the trending ones were developed in China or derived from a Chinese model.
• In one year, the most-liked models went from mainly Meta's Llama family to an international mix with DeepSeek-R1 at the top.
• Alibaba has more derivative models than Google and Meta combined. Its Qwen family alone has more than 113,000 fine-tunes, merges and other adaptations.

One caveat from the same report: models by individual users and by organisations without a clear home country account for about half of all downloads.

What changed in China after DeepSeek-R1?

DeepSeek released R1 in January 2025, a reasoning model that the company says performs comparably to OpenAI's o1 on maths, code and reasoning. The weights came under the MIT licence, which allows commercial use and distillation into other models. A month earlier, DeepSeek had reported that the full training of its V3 base model took 2.788 million H800 GPU hours (DeepSeek-V3 technical report).

According to Hugging Face, China's AI ecosystem "steered heavily into open source" after R1. Baidu went from zero releases on the Hub in 2024 to more than 100 in 2025. ByteDance and Tencent each released eight to nine times as much as before. Baidu and MiniMax, which had favoured closed models, switched to open releases.

Release pace counts as well. Interest in an open model peaks right after release and lasts about six weeks on average, and DeepSeek stayed relevant through a series of releases (V3, R1, V3.2). Alibaba's Qwen3.5 family (February 2026) shows the breadth: eight open-weight models from 0.8 to 397 billion parameters under the Apache 2.0 licence. In Alibaba's own tests, reported by The Batch, Qwen3.5-9B beat OpenAI's gpt-oss-120b, a model more than ten times its size, on most language benchmarks apart from reasoning and coding.

Why does hardware matter here?

On 20 March 2026, The Batch reported, citing Reuters, that DeepSeek had not yet given Nvidia or AMD early access to its upcoming DeepSeek-V4, as is usual before a major release, but had shared a prerelease version with Huawei. That gave the Chinese chipmaker several weeks to tune its software for the model. DeepSeek did not explain the decision. The Batch called the move possibly "more symbolic than significant" and noted that it fits China's goal of technological self-sufficiency.

Hugging Face sees the same direction: Chinese open models are released with explicit support for chips made in China, and Alibaba invests in inference chips for Chinese data centres.

The role of US export controls is mixed. The Batch argues that the chip restrictions "largely backfired by spurring China to build its domestic chip industry". Yet an unnamed senior US official told Reuters that DeepSeek-V4 was trained on Nvidia's most advanced chips despite those controls.

Where else is open source growing?

Robotics is one of the fastest-growing communities on the Hub. The number of robotics datasets went from 1,145 in 2024 to 26,991 in 2025, which made robotics the largest dataset category. Text generation came second with about 5,000. Hugging Face bought the open-source robot maker Pollen Robotics in April 2025, and the GitHub stars of its robotics library LeRobot nearly tripled in a year. The report does not break the robotics growth down by country.

Independent developers gained ground too. Their share of downloads rose from 17% before 2022 to 39% in 2025, while the share of industry fell from about 70% to 37%. Most of them quantise, adapt and redistribute base models, so in Hugging Face's words they "steer a meaningful portion of what typical users can run".

How is the West responding?

Hugging Face sees Western organisations looking for commercially usable alternatives to Chinese models. That puts weight on open-weight projects such as OpenAI's gpt-oss, Ai2's OLMo and Google's Gemma. Whether they can match the adoption of Qwen and DeepSeek will be "a defining question of 2026", the report says.

Open-source investment also goes into tooling. On 19 March 2026, OpenAI announced that it will acquire Astral, the maker of the Python tools uv, Ruff and ty. After the deal closes, which requires regulatory approval, the Astral team joins OpenAI's Codex team. Both companies say OpenAI will keep supporting the open-source tools (Astral).

What should developers take from this?

• Test Chinese open models on your own tasks. Qwen and DeepSeek lead adoption on the Hub, and Qwen3.5 covers sizes from 0.8 to 397 billion parameters.
• Read the licence of the exact model. DeepSeek-R1 uses MIT and Qwen3.5 uses Apache 2.0, but a derivative can carry another licence: DeepSeek's own distilled models built on Llama fall under Meta's Llama licence.
• Check security before you build agents on a model. In September 2025 CAISI, the AI evaluation centre of the US standards institute NIST, found that agents based on DeepSeek R1-0528 were on average 12 times more likely than US frontier models to follow planted instructions that hijack them (see prompt injection). It also found that DeepSeek's models echoed four times as many misleading Chinese Communist Party narratives as US reference models.
• Expect small models to do most of the work. The median size of a downloaded model on the Hub grew only from 326 million to 406 million parameters between 2023 and 2025, while the mean rose from 827 million to 20.8 billion.