欧

Alibaba's open-weight "Qwen" (Tongyi Qianwen) AI model family has become the world's most-downloaded open-source AI model family, overtaking US tech giants Meta and Google to claim the top spot in open-source AI downloads. According to a "State of Open Models" report published on August 14 by Hugging Face — the world's leading platform for hosting and downloading AI models — and reported by Bloomberg, Fortune and other international outlets, Qwen's cumulative global downloads surpassed 3 billion over the past roughly six months, marking a new milestone for the global reach of China's open-source AI ecosystem.

By comparison, the report shows that over roughly the same period, Google's models recorded about 418 million downloads and Meta's about 227 million — combined, still far short of Qwen's more than 3 billion downloads. That means Qwen has opened up a substantial lead over the two US tech giants on this key metric of open-model downloads, becoming one of the go-to base models global developers choose to build on and fine-tune. It should be noted that download counts are not identical to unique user numbers, since the same model is often downloaded repeatedly, deployed across different environments, or used to create derivative versions — but as a core measure of an open model's global reach, the figure still carries significant weight.

Echoing the rise in downloads, Chinese large language models' share of global AI "usage" — measured in compute — is also continuing to grow. According to data from AI model-routing platform OpenRouter, in the week of August 3-9, global AI model token usage totaled about 69 trillion tokens, of which Chinese-made models accounted for about 34.25 trillion tokens combined, up roughly 21.76% from the previous week. This marked the 15th consecutive week in which Chinese models' weekly token usage exceeded that of US models, reflecting that global developers and enterprises are increasingly relying on Chinese models to actually carry out tasks — not merely to download and try them out.

This trend is not a short-term blip. Earlier OpenRouter data showed that as of May this year, Chinese open-weight models already accounted for about 61% of all token consumption on the platform, with four of the five most-used models built in China, while Meta's Llama family — which dominated the open-model field two years ago — had fallen out of the leading rankings. In the model-usage rankings for the week of August 3-9 alone, a DeepSeek model topped the list with about 9.39 trillion tokens of usage, followed closely by a Tencent model with about 8.94 trillion tokens. Combined with other Chinese models already ranked highly, such as Moonshot AI's Kimi and Zhipu's GLM, this shows that Chinese-made models continue to gain more top spots in global usage rankings, rather than being driven by any single model alone.

Public information shows Alibaba has released a steady stream of open-weight Qwen models since 2023, open-sourcing more than 460 versions of varying sizes and purposes to date, spanning language models ranging from a few billion to several hundred billion parameters, most released under relatively permissive open-source licenses that let global developers freely download, deploy and further develop them. On this foundation, global developers have produced roughly 300,000 derivative models fine-tuned or retrained from Qwen. This year Alibaba has also rolled out new versions in quick succession, including the flagship Qwen3.6-Plus released mid-year, and new models such as Qwen3.8-2.4T-A95B and Qwen3.8-27B, open-sourced on August 12 and 14 respectively, continuing to expand its model lineup. Hugging Face's report described Qwen as having "become one of the largest foundations of the open AI ecosystem," adding that it has "become part of the default workflow for developers deciding what models to fine-tune and deploy."

Analysts say Qwen's ability to achieve global adoption on this scale is closely tied to differentiated technical choices. Unlike earlier versions of Meta's Llama, which were mainly optimized for English, the Qwen family emphasized multilingual capability from its early versions onward, with the latest generation reportedly supporting more than 200 languages. At the same time, Qwen's tokenizer is markedly more efficient at processing non-English text such as Chinese — for Chinese text, Qwen's tokenizer can encode at roughly one token per character, whereas the byte-pair encoding used by Llama 3 often needs 3 to 5 tokens to represent a single Chinese character. For global enterprise users based mainly outside English-speaking markets, this means the computational cost of running inference and deployment with Qwen can be significantly lower than with comparable Western models. On performance, third-party evaluations show that this year's Qwen 3.5 series, on coding benchmark SWE-bench Verified and reasoning benchmarks such as GPQA Diamond and AIME, performs on par with or even better than some mainstream international models released around the same time, while its size range — from lightweight versions with a few billion parameters to flagship versions with nearly a hundred billion — lets enterprises with varying compute budgets choose as needed.

More broadly, Qwen's rise to the top of the download charts and Chinese models' dominance of global token usage are emblematic of a wider wave over the past two years in which Chinese AI companies have released a dense succession of high-performing open-weight models. Since early 2025, when DeepSeek released its R1 model and demonstrated that a Chinese lab could match top Western closed-source models on reasoning ability while simultaneously open-sourcing its weights, companies including Alibaba, Tencent, Moonshot AI (Kimi), Zhipu (GLM) and MiniMax have successively released their own open models, filling in nearly every price-and-performance tier from entry-level up to near the international frontier. By contrast, some leading Western labs, including OpenAI, Anthropic and Google, still largely operate under closed-source or restrictively licensed commercial models, creating a sharp contrast between the two approaches. Several industry observers have suggested this divide reflects both differing commercial strategies among labs and Chinese companies' interest in using open-source releases to rapidly expand their global developer base and gain a greater say in setting AI infrastructure standards.

Notably, Alibaba itself is also exploring a balance between open and closed approaches: according to reports, the company has shifted its most powerful flagship models, Qwen3.7-Max and Qwen3.7-Plus, to closed-source status, offering them only through paid APIs — putting Alibaba in more direct competition with companies like OpenAI and Anthropic in the enterprise API market. At the same time, Alibaba is reportedly testing a revenue-sharing mechanism for some of its open flagship models, such as the open-weight version of Qwen3.8-Max. This suggests that while continuing to use an open-source strategy to expand its global ecosystem influence, Alibaba is also exploring new commercialization paths for its most advanced models.

Industry observers say China's twin lead in open-model downloads and usage is reshaping global patterns of AI infrastructure adoption: for many developing markets, multilingual use cases and cost-sensitive enterprises, competitively performing open models that can be freely downloaded, modified and deployed locally are becoming an increasingly attractive alternative to closed commercial APIs, while also helping reduce countries' and companies' reliance on a single vendor. As Alibaba and other Chinese firms continue updating their model lineups, expanding their developer ecosystems, and exploring commercial models that combine open and closed approaches, whether China's open-source AI models can maintain this lead will be an important bellwether for how the global AI industry landscape evolves.