Qwen is Alibaba Cloud’s family of large language models — and, by the numbers that matter for an ecosystem, the most consequential open-source AI family in the world. As of Hugging Face’s “State of Open Source” report (Spring 2026), Qwen has spawned more than 113,000 derivative models — around 200,000 if you count everything tagged Qwen — more than Google and Meta combined. Cumulative downloads passed 700 million by January 2026 (Forbes, May 2026), helping push Chinese-developed models to 41% of all Hugging Face downloads in 2025. On OpenRouter, the neutral routing layer for AI usage, Qwen handles 13.9% of all routed tokens — roughly 2.77 trillion tokens a week — behind only DeepSeek among individual vendors (CNBC analysis of OpenRouter data, 7 July 2026). And since relaunching its consumer app in November 2025, Alibaba has built a 167 million monthly-active-user assistant, up roughly 58-fold year on year (QuestMobile, May 2026 data). The strategy is best understood as the Android of AI: give the operating layer away, become the default everyone builds on, and monetise the infrastructure underneath. This profile looks at how Alibaba got here, who builds on Qwen, and what could still derail it.
Where did Qwen come from?
Qwen began as Tongyi Qianwen, a chatbot Alibaba Cloud launched in beta in April 2023 and opened to the Chinese public that September after regulatory clearance. The consequential decision came in August 2023, when the team released Qwen-7B’s weights openly — the start of a cadence that has barely paused since. Development sat inside Alibaba Cloud’s Tongyi Lab, with Lin Junyang leading the Qwen model team and Zhou Jingren as chief AI architect; Lin reportedly departed in March 2026 shortly after shipping Qwen3.5, and Alibaba consolidated its AI operations into a “Token Hub” unit under group CEO Eddie Wu in April 2026.
The recent model cadence has been relentless: Qwen3 arrived under an Apache 2.0 licence in April 2025, Qwen3.5 followed as the open flagship on 16 February 2026, and the Qwen3.6 series landed in April 2026. Then, on 19 July 2026 at the World AI Conference in Shanghai, Alibaba previewed Qwen3.8-Max — a 2.4 trillion-parameter sparse mixture-of-experts model, its first multimodal model above one trillion parameters, announced two days after Moonshot AI’s 2.8 trillion-parameter open-weight Kimi K3 (MarkTechPost, 19 July 2026). Alibaba claims it ranks second only to Anthropic’s Fable 5 on internal evaluations, though as of 25 July 2026 no benchmark table, model card or licence has been published, and the promised open weights have no date (TechTimes, 21 July 2026).
Why does Alibaba give its best models away?
Because the models are the funnel, not the product. Open weights make Qwen the default base layer for developers, startups and enterprises — and when those workloads need serious inference, fine-tuning or agentic infrastructure, Alibaba Cloud is the natural home. The pull-through is now visible in the accounts. Alibaba reported its 11th consecutive quarter of triple-digit AI-related revenue growth in the March 2026 quarter, is targeting 30 billion yuan (US$4.4 billion) in annualised AI recurring revenue by the end of 2026, and expects AI to exceed 50% of cloud-computing revenue within a year (SCMP, 13 May 2026). Cloud revenue grew around 38% year on year in that quarter (CNBC, 13 May 2026).
The capital commitment matches the ambition. Alibaba pledged 380 billion yuan (~US$53 billion) to AI and cloud infrastructure over three years in February 2025 — and Eddie Wu said in May 2026 the company would likely “overshoot” that figure, calling the original number “on the small side” (SCMP; Reuters, May 2026). Crucially, this is not speculative buildout: Wu told investors in late 2025 that “all of our GPUs are running at full capacity” and Alibaba was rationing access because supply cannot keep up with demand — a constraint he expected to last two to three years (Data Center Dynamics, November 2025). The Android analogy holds: Google gave away the OS and monetised services and search; Alibaba gives away the weights and monetises the tokens. For the wider funding picture, see our China AI market overview.
The 113,000-derivative ecosystem
Hugging Face’s Spring 2026 report is the clearest measure of Qwen’s gravitational pull: 113,000+ derivative models — fine-tunes, quantisations, merges and distillations — more than Google and Meta combined, on a platform that reached 13 million users and two million public models in 2025. The same report found Chinese models overtook US models in monthly downloads during 2025.
Who actually builds on Qwen? Almost everyone who needs a capable, permissively licensed base. Startups fine-tune Qwen because Apache 2.0 allows commercial use without copyleft strings. Enterprises deploy it because it is, in Airbnb CEO Brian Chesky’s words, “very good… also fast and cheap” (Bloomberg, October 2025). National and corporate “sovereign” model programmes across Asia have quietly used Qwen as scaffolding — a pattern we track in our guide to sovereign AI programmes across Asia. Usage follows: Chinese-origin models peaked at 46% of enterprise token volume on OpenRouter by mid-2026, against 35.7% for all US models combined, while pricing 60–90% below US equivalents (CNBC, 7 July 2026). The economics do a lot of the evangelising.
From research lab to 167M-user consumer app
For two years Qwen was a developer brand with no consumer footprint to speak of — Alibaba’s answer to ByteDance’s Doubao was conspicuously absent. That changed in mid-November 2025, when Alibaba relaunched the Qwen app as a public beta positioned as “the best personal AI assistant,” with deep research, image generation and slide creation built in. Within weeks it became the world’s fastest-growing AI app, with MAUs up 149% in November 2025 to 18.34 million (SCMP, 3 December 2025). By May 2026 it had reached 167 million MAU — a roughly 58-fold (+5,792.9%) year-on-year increase (QuestMobile data, July 2026) — making it China’s second-largest AI-native app behind Doubao’s 382 million and ahead of DeepSeek’s 130 million, in a domestic market that hit 499 million AI-native app users, up 85.4% year on year (QuestMobile via TechNode, 14 July 2026).
The relaunch closes Alibaba’s loop: the app plugs Qwen into commerce, food delivery and payments across the Alibaba ecosystem, turning model quality into consumer lock-in the way Android turned OS share into Google service usage — a third leg alongside open weights for developers and cloud APIs for enterprises.
Where Qwen shows up that you don’t expect
The depth of Qwen’s reach keeps surfacing in unexpected places. Apple Intelligence in mainland China runs on Alibaba’s models, under a partnership first reported in February 2025 — meaning Qwen sits inside Chinese iPhones. Airbnb’s customer-service AI leans heavily on Qwen rather than OpenAI in production (Bloomberg, October 2025). And in January 2026, South Korea delivered the most telling example: Naver — the country’s flagship internet company — was cut from the government’s $1 billion sovereign AI foundation-model programme after evaluators found its HyperCLOVA X SEED model’s vision encoder showed 99.5% cosine similarity with Qwen 2.5-VL’s, accounting for about 12% of the model’s parameters (Bloomberg, 15 January 2026; byteiota, January 2026). A programme designed to reduce dependence on foreign AI discovered that its national champion had built on Chinese open weights. LG, SK Telecom and Upstage advanced instead (Korea Herald, January 2026). Similar dynamics run through Southeast Asia, where regional models and startups increasingly build on Chinese bases — see our analysis of Chinese AI models across Southeast Asia.
What could go wrong?
US scrutiny. In April 2026, the House committees on China and homeland security wrote to Airbnb demanding answers on its Qwen deployment, with China committee chair John Moolenaar warning that Chinese models are “trained by China’s censorship regime and introduce hidden vulnerabilities” (Forbes, 21 May 2026). Chesky called it a “misunderstanding” — self-hosted open weights send no data to China — but formal restrictions on Chinese model weights, or procurement bans that taint Qwen derivatives, remain the ecosystem’s biggest tail risk. Our US vs China AI capability tracker follows this in detail.
Competition on both flanks. DeepSeek still routes more tokens than Qwen on OpenRouter (17.6% vs 13.9%, CNBC, July 2026) and remains the most-liked model on Hugging Face; Moonshot’s Kimi K3 open-weighted 2.8 trillion parameters before Alibaba did. At home, Doubao’s 382 million MAU is more than double Qwen’s. Execution risk is real too: margins are being squeezed by capex, the flagship Max tier is drifting closed — Qwen3.8-Max shipped with no benchmarks, card or licence — and the reported March 2026 departure of model lead Lin Junyang tests the team’s depth. The Android playbook only works if the free layer stays best-in-class.
Frequently asked questions
Is Qwen really open source?
Mostly. Core model families — Qwen3, Qwen3.5 and the open Qwen3.6 sizes — ship weights under Apache 2.0, permitting free commercial use. But the top “Max” and “Plus” tiers (Qwen2.5-Max, Qwen3.6-Plus, Qwen3.8-Max) are proprietary API products, so “open-weight with a closed frontier” is the accurate description as of July 2026.
How does Alibaba make money from a free model?
Cloud pull-through. Open Qwen adoption drives inference, fine-tuning and agent workloads to Alibaba Cloud, which logged 11 straight quarters of triple-digit AI revenue growth to March 2026 and targets US$4.4 billion in annualised AI recurring revenue by end-2026 (SCMP, May 2026).
Is Qwen bigger than Llama now?
On ecosystem measures, yes. Qwen’s 113,000+ derivatives exceed Google and Meta combined on Hugging Face (Spring 2026 report), and Meta’s Llama has fallen below 1% of OpenRouter routed volume by mid-2026 (Data Gravity analysis, 2026).
Is it safe for non-Chinese companies to build on Qwen?
Technically, self-hosted weights transmit nothing to Alibaba. The live risks are political and regulatory: US congressional scrutiny of Qwen users began in April 2026, and sovereign programmes such as Korea’s now audit for Chinese model lineage — as Naver learned when it was cut in January 2026.
What is Qwen3.8-Max?
Alibaba’s largest model to date: a 2.4 trillion-parameter multimodal mixture-of-experts model previewed at WAIC Shanghai on 19 July 2026, handling text, images, video and documents. Alibaba says open weights are “coming soon”; as of 25 July 2026, no benchmarks, licence or release date have been published.
For the regional context around Qwen’s rise, start with our Asia AI market hub.
Discover more from Digital in Asia
Subscribe to get the latest posts sent to your email.
