In February 2026, Alibaba’s Qwen model exceeded on Hugging Face the sum of the eight main competitors combined in terms of monthly downloads, representing more than 50% of global downloads. Qwen had already crossed the one billion cumulative downloads mark on January 21, 2026. To put the scale of this shift in perspective: Chinese models’ share of global open-source AI was below 2% at the end of 2024. It reached nearly 30% in certain peak weeks at the end of 2025, according to OpenRouter data compiled by Andreessen Horowitz — though the annual average remained around 13% of token volume.

This is not an industrial accident. It is the deployment of an explicit strategy, whose logic resembles the path Android followed in 2008: distribute a free operating system to impose a global norm, then capture value in the ecosystem built around it.

The Essentials

  • Alibaba’s Qwen had already surpassed Meta’s Llama in cumulative total by October 2025, and in monthly downloads by December 2025. The one billion cumulative downloads mark was crossed on January 21, 2026.
  • China’s share in global open-source models increased from 1.2% at the end of 2024 to nearly 30% in peak weeks at the end of 2025 (annual average of approximately 13%), according to OpenRouter data analyzed by Andreessen Horowitz
  • Moonshot AI’s Kimi K2.5 costs four times less than GPT-5.2 at equivalent performance, according to Artificial Analysis
  • Zhipu AI’s GLM-5.2 (Z.ai), released on June 13, 2026 under an MIT license, positions itself as the first open-weight model with performance comparable to the best closed models for code and agent tasks — at approximately $1.40 per million input tokens versus $5 for GPT-5.5
  • Meituan’s LongCat-2.0, released on June 30, 2026, is a 1.6 trillion parameter model trained entirely on Chinese domestic chips — without any Nvidia chips — on a cluster of 50,000 accelerators
  • The challenge for Europe is not to choose a side, but to build the institutional capacity to evaluate, adapt, and secure models it did not train

Qwen Surpasses Llama, and This Is Just the Beginning

For a long time, the hierarchy of open models seemed stabilized. Meta published Llama, the global community adopted it, and developers built on it. This configuration had a logic: Meta had the resources to train competitive models, and trust in a known American player reduced friction to adoption.

Qwen changed the equation. Alibaba published a succession of versions since 2023, each more performant and cheaper to run than the previous one. Qwen2.5 and specialized variants for code, mathematics, and reasoning progressively covered the most important use cases. The result is visible in benchmarks: according to Artificial Analysis data, several Qwen variants regularly appear in the top ten open models, across all nationalities.

Progress is not unique to Alibaba. DeepSeek produced in January 2025 a reasoning model that rivals the best American closed models at a fraction of their training cost. Moonshot AI, with Kimi K2.5, offers performance comparable to GPT-5.2 at a quarter of its inference price. Baidu, Zhipu, 01.AI follow their own publication pace. What appears to be a sudden convergence is actually the result of several years of coordinated investment, a dense research community, and intense internal competitive pressure.

Two recent releases illustrate how far this dynamic has been pushed. GLM-5.2 is Zhipu AI’s (Z.ai) latest flagship model — designed specifically for code, reasoning, and agentic tasks — released on June 13, 2026 under an MIT open-source license. It is a mixture-of-experts model with 744 billion total parameters and 40 billion active parameters, with a quadrupled context window at one million tokens. On Artificial Analysis’ Intelligence Index v4.1, GLM-5.2 scores 51, ahead of MiniMax-M3 (44), DeepSeek V4 Pro (44), and Kimi K2.6 (43). Via providers like OpenRouter, the model costs approximately $1.40 per million input tokens and $4.40 per output, compared to $5/$30 for GPT-5.5 and $5/$25 for Claude Opus.

LongCat-2.0, released by Meituan, is a MoE model with 1.6 trillion parameters with approximately 48 billion active parameters per token, trained on over 30 trillion tokens, designed for agentic code. It supports a context window of one million tokens, and Meituan claims that the entire training and large-scale deployment was accomplished on a cluster of 50,000 Chinese domestic chips. This detail deserves examination. Where DeepSeek-V4-pro only used domestic chips for inference — the least demanding task, responding to requests — LongCat-2.0 used domestic hardware for both inference and pre-training. This is a qualitative break in demonstrating hardware sovereignty.

The Android Logic Applied to AI

Google opened Android in 2008 not out of philanthropy, but to prevent Microsoft or Apple from controlling the mobile software layer. By making the system free and open, Google ensured that the application ecosystem would be built on its standards, its APIs, its identity architecture. Value is not captured in the license: it is captured in the services that run on it, in the data generated, in the usage habits that make replacement expensive.

Chinese strategy in open-source AI follows the same architecture. Publishing Qwen for free ensures that developers worldwide learn to work with its parameters, its file formats, its architectural conventions. Companies building products on Qwen develop technical and organizational dependencies. When an improved version is released, they have reasons to stay in the ecosystem rather than start over on another basis.

The comparison has a limit: Google retained control of Android’s core and associated services. In open-source AI, Alibaba publishes the model’s complete weights, allowing theoretically anyone to duplicate it, modify it, run it without ever returning to Alibaba. This more complete openness serves a different objective: normalizing the architecture rather than service loyalty. When a model becomes the reference on which developer training, online tutorials, university courses, and evaluation tools are built, its footprint transcends the platform distributing it.

The LongCat-2.0 case adds a dimension the Android metaphor did not anticipate: the hardware question. Meituan’s announcement puts pressure on one of the simplest assumptions underlying American export control policy — namely, that depriving Chinese actors of the latest Nvidia chips would prevent them from training and deploying frontier-level systems at massive scale. If software freedom was already a tool for capturing norms, demonstrating complete hardware autonomy changes the nature of the power dynamic.

MIT Technology Review notes that the U.S. Commission on China (USCC) has begun explicitly analyzing this strategy as a vector of influence over international AI standards, just as patents in telecommunications or industrial standards in manufacturing have been.

A Planet That Cannot Afford to Pay for Claude

The geography of adoption is perhaps the most structuring element of this dynamic. OpenAI’s GPT-5.2 and Anthropic’s Claude Sonnet are exceptional products in their category. They are also structurally inaccessible to the majority of developers and companies in middle or low-income countries.

A developer in Lagos, Karachi, or Bogotá wanting to build an AI product has very different options depending on their budget. Closed American models involve API costs that, at scale, become prohibitive for a local market. Chinese open models, by contrast, can be deployed locally on modest hardware, without recurring inference fees, with documentation available in multiple languages.

This price differential is not marginal. Kimi K2.5 at a quarter the price of GPT-5.2 means concretely that a developer can afford four times more experimentation, four times more iterations, four times more end users before hitting their budget ceiling. At standard rates, LongCat-2.0 is 6 to 7 times cheaper per token than GPT-5.5 on input, and 10 times cheaper on output. For a startup in Jakarta or an administration in Nairobi, this is often the difference between a viable project and an abandoned one.

The geopolitical context sharpens this imbalance. The Trump administration ordered the blocking of Anthropic’s most advanced models — Fable 5 and Mythos 5 — for foreign nationals, and it was the same day that Zhipu announced the open-source release of GLM-5.2 without any usage restrictions. For developers outside the United States, GLM-5.2 is therefore the most performant open-license model currently available. The window left open by American restriction policy, Chinese models occupy methodically.

The geopolitical stakes are direct: actors who shape the digital tools of middle-income countries also progressively shape development habits, data formats, integration practices. This is cognitive infrastructure, not just technology.

What Freedom Actually Conceals

It would be naive to read this strategy without its blind spots. Freedom is not neutral, and questions of trust in distributed AI models are legitimate at several levels.

The first level is technical. An open-source model whose weights you possess can be audited, modified, and deployed without connection to external servers. This is a real security property, distinct from closed models where the operator remains in the loop. But auditing a model with tens of billions of parameters is a task beyond the reach of most organizations adopting it. Formal audit capacity and actual audit practice are two very different things.

The second level is regulatory. Models trained in China are subject to Chinese regulations on content and information, which include obligations to filter certain subjects. These filters are not necessarily visible in technical benchmarks, which measure performance on neutral tasks. An organization deploying these models in sensitive contexts must question what the model refuses to do and why.

The third level is usage data. Unlike a closed API, a local deployment of an open model does not generate calls back to the creator’s servers. But tools for fine-tuning, evaluation platforms, extensions, and plugins gravitating around ecosystems can collect data. Anyone using Z.ai’s cloud API is subject to Chinese law — a constraint that disappears with MIT-licensed self-hosting of weights. The boundary between open model and data collection infrastructure is not always clear.

These risks exist. They do not justify precautionary exclusion. They justify serious evaluation, which is different.

What Europe Can Build — and Has Not Yet

Europe produces few competitive large models. Mistral is the notable exception, and its trajectory illustrates both what is possible and what is structurally lacking. The absence of European technology giants is not a natural inevitability: it is the result of accumulated choices in regulation, funding, and risk tolerance. In foundational AI, the dice are cast for this generation of models. Europe will not have a competing GPT-5 within the next eighteen months.

What Europe can build is institutional competence in evaluation and adaptation. The AI Act creates a compliance obligation for models deployed on the European market, regardless of origin. This obligation is also an opportunity: it forces development of audit methodologies, competent bodies, evaluation standards. These standards, if well designed, could become a global reference, as certain European standards have been in chemistry, automobiles, or data protection.

The realistic trajectory for European institutions is not to produce models capable of rivaling the best American or Chinese ones on all benchmarks. It is to build the capacity to take any open model, understand it, adapt it to local context, secure it for sensitive uses, and evaluate it according to transparent criteria. This is an integrator and regulator competence, not a base creator one.

This position is not comfortable. It supposes accepting partial technological dependence, managing risks one did not entirely choose, and building value on foundations one does not fully control. But the question of AI agent governance arises in all cases, including for American models: no one in Europe controls GPT-5.2’s weights either.

Institutional competence to evaluate and adapt has an added virtue: it is transferable. A European state or company that knows how to audit a Chinese open model also knows how to audit an American model, a local model, a competing model. This is a general capacity, not a specific dependence.

What Developers Vote for With Their Keyboards

Institutional benchmarks and geopolitical reports have their logic. But the most direct signal remains downloads. When a developer chooses Qwen over Llama to build their application, they are not necessarily making a political choice. They are making a pragmatic choice: performance, inference cost, deployment ease, documentation quality. These factors are measurable, comparable, and for now, they often tip toward Chinese models on segments where cost matters.

The LongCat-2.0 case illustrates how far this dynamic can go without even claiming to do so. For two months before its public revelation, the model had circulated anonymously on OpenRouter under the alias “Owl Alpha,” accumulating approximately 10.1 trillion tokens monthly — an explosion of 242% month-on-month — propelling it into the global top three on the platform. This anonymous model had ranked first on Hermes Agent, second on Claude Code, and third on OpenClaw in call volume. Developers had chosen a model whose origin they did not know. Pragmatism was their only filter.

This reality will not disappear with calls for geopolitical vigilance. It calls for a concrete industrial and institutional response. On the American side, restrictions on chip exports continue to constrain Chinese training capacities, even if their effectiveness is partial and contested. LongCat-2.0 does not invalidate the logic behind American export controls — restrictions still increase costs, slow access, complicate scaling, and force Chinese companies into harder engineering compromises. But they are no longer sufficient to prevent production of frontier-level models. On the European side, the challenge is to quickly build the evaluation bodies and standards the AI Act makes necessary, before de facto norms are fixed by massive adoption practices.

The Android strategy took ten years to produce its full effects. The Qwen strategy is perhaps halfway there. The window to build a position as a credible regulator and competent integrator is open. It is not open indefinitely.

Sources

  1. MIT Technology Review, “What’s next for Chinese open-source AI?” (February 2026): https://www.technologyreview.com/2026/02/12/1132811/whats-next-for-chinese-open-source-ai/
  2. Andreessen Horowitz (a16z), OpenRouter data on open-source model market share (2025-2026)
  3. Artificial Analysis, performance and cost comparisons of inference models (2026)
  4. US-China Economic and Security Review Commission (USCC), 2025 annual report, AI and international standards section
  5. Hugging Face, model download statistics (public access, March 2026)
  6. Trending Topics / Eigent.ai, analysis of Zhipu AI’s GLM-5.2 (Z.ai), June 2026: https://www.eigent.ai/blog/glm-5-2
  7. VentureBeat, “Meituan open sources LongCat-2.0” (June 30, 2026): https://venturebeat.com/technology/meituan-open-sources-longcat-2-0
  8. Decrypt, “LongCat-2.0: The Stealth AI Model That Was Quietly Topping OpenRouter All Along” (July 2026): https://decrypt.co/372579/longcat-2-0-meituan-ai-stealth-model-openrouter
  9. South China Morning Post, “China claims biggest AI model trained on local chips, as Meituan releases LongCat-2.0” (July 2026): https://www.scmp.com/tech/tech-trends/article/3358854/china-debuts-biggest-ai-model-trained-local-chips-meituan-releases-longcat-20
  10. OpenRouter State of AI 2025 (primary source): https://openrouter.ai/state-of-ai
  11. South China Morning Post — Qwen 50% global downloads: https://www.scmp.com/tech/big-tech/article/3349552/alibabas-qwen-family-captures-over-50-global-open-source-downloads-report-finds
  12. Gizmochina — Qwen ~942M downloads (March 2026): https://www.gizmochina.com/2026/04/10/alibaba-qwen-ai-downloads-1-billion-leads-open-source-ai/
  13. Pandaily — Qwen 1 billion downloads (January 21, 2026): https://pandaily.com/alibaba-s-qwen-open-source-models-surpass-1-billion-downloads-ranking-first-globally
  14. Trending Topics — GLM-5.2 full coverage: https://www.trendingtopics.eu/glm-5-2-chinas-zhipu-ai-beats-even-googles-top-models-with-its-new-open-llm/
  15. DataNorth — GLM-5.2 official specs: https://datanorth.ai/news/zhipu-ai-releases-glm-5-2
  16. Axios — Anthropic Fable 5 / Mythos 5 blocking: https://www.axios.com/2026/06/12/anthropic-trump-mythos-fable-national-security
  17. Epoch AI — DeepSeek R1 January 2025: https://epoch.ai/gradient-updates/what-went-into-training-deepseek-r1
  18. OpenRouter — GPT-5.5 pricing: https://openrouter.ai/openai/gpt-5.5
  19. Anthropic official — Claude Opus pricing: https://www.anthropic.com/claude/opus