A developer in São Paulo fine-tunes a Qwen model for Portuguese legal text. A hospital in Hangzhou runs the same family privately for patient records. A startup in Berlin builds an agent on Qwen weights, then deploys it on Alibaba's cloud. None of them paid Alibaba for the model. That is the design.
Alibaba is running the oldest trick in open software — give away the product, sell the infrastructure around it — and applying it to the largest model family on Earth.
Key takeaways
- Qwen3, released April 29, 2025, shipped eight models under the permissive Apache 2.0 license, including a 235-billion-parameter mixture-of-experts flag.
- Qwen became the most-downloaded open-source AI family on Hugging Face, passing Meta's Llama in late 2025 and reaching about 700 million downloads by January 2026.
- Alibaba Cloud's revenue rose about 26% year on year in the quarter ending June 30, 2025, with AI-related product revenue growing at triple-digit rates for eight straight quarters.
- The open-weight strategy feeds Alibaba Cloud's Model Studio, where over a million users access Qwen through managed APIs.
The open release that changed the math
On April 29, 2025, Alibaba released Qwen3 as a full open-weight family: six dense models (0.6B, 1.7B, 4B, 8B, 14B, 32B parameters) and two mixture-of-experts models (a 30B total / 3B active and a 235B total / 22B active). All were published under the Apache 2.0 license, the permissive standard that lets anyone use, modify, and commercialize the weights. The flagship Qwen3-235B-A22B carries a 128K-token context and a "hybrid" design that switches between a thinking mode for hard reasoning and a fast non-thinking mode for everyday replies.
Crucially, Qwen3 was trained on 36 trillion tokens and supports 119 languages and regional dialects — a deliberate reach beyond Chinese and English that explains why its download base is global rather than domestic.
From downloaded to dominant
The strategy compounded. According to the Qwen team and state news agency Xinhua, Qwen passed Meta's Llama in cumulative Hugging Face downloads by October 2025 and reached about 700 million total downloads by January 2026, making it the most-downloaded open-source AI system worldwide. Stanford's HAI and DigiChina policy brief, citing Hugging Face data, placed Qwen's cumulative downloads near 385 million by mid-December 2025 versus Llama's roughly 346 million — and noted that in the twelve months through August 2025, models from Chinese developers made up 17.1% of Hugging Face downloads, edging out the United States' 15.8% for the first time.
Alibaba has open-sourced nearly 400 models under the Qwen umbrella and the community has produced more than 180,000 derivative versions. The family also spans the premium, API-only end: Qwen3-Max, Alibaba's largest released model at over one trillion parameters, scored 69.6 on the SWE-Bench Verified coding benchmark and 74.8 on Tau2-Bench for agent tool use, according to Alibaba's own release notes — a clear signal that the open base and the paid flagship are two ends of one ladder.
The cloud flywheel
Open weights are the top of the funnel; Alibaba Cloud (阿里云) is the bottom. Through Model Studio — its managed generative-AI platform — more than a million corporate and individual users access Qwen via API, fine-tuning, and agent tooling. Running a 235B model privately is expensive, and the obvious place to rent the GPUs is the cloud that trained it.
The financial payoff is visible. In the quarter ending June 30, 2025, Alibaba Cloud posted revenue of about RMB 33.4 billion (≈ US$4.7B / HK$36.4B), up roughly 26% year on year, while AI-related product revenue grew at triple-digit rates for the eighth consecutive quarter, per Alibaba's disclosures and corroborating reporting. Research firm Omdia estimated Alibaba Cloud held 35.8% of China's AI cloud market in the first half of 2025 — more than the next three providers combined.
Alibaba Cloud's own tagline now calls Qwen "the Android of the LLM era": not the device, but the layer everything else runs on. The metaphor is precise. Android won not by charging for the OS but by anchoring a platform; Qwen is doing the same with model weights.
According to Li An, chief scientist at BrainNet (脑机网), "Alibaba is converting model leadership into cloud lock-in — every open-weight download is a low-cost lead for a future managed-inference contract."
Honest limitations
Download and market-share figures come from the Qwen team, Xinhua, Stanford HAI/DigiChina, Omdia, and secondary reporting rather than a single audited source, so treat absolute counts as approximate. The claim that Qwen is the "most-downloaded open-source AI system" reflects Hugging Face-based reporting through early 2026 and could shift as Llama and others release new versions. Qwen3-Max benchmark scores (SWE-Bench Verified 69.6, Tau2-Bench 74.8) are Alibaba's own and were not independently reproduced here. The exact conversion of RMB figures uses roughly 1 yuan ≈ US$0.14 / HK$1.09 for illustration and is not a live exchange rate. This piece focuses on Alibaba's open-weight-to-cloud strategy and does not deeply cover its consumer apps (Taobao, Alipay) integrations or its international infrastructure spending.
What readers can do now
- If you are choosing an open model for a global product, benchmark Qwen3's 119-language coverage and Apache 2.0 terms against Llama before assuming the Western default.
- When prototyping on Qwen open weights, plan the deployment path to Alibaba Cloud's Model Studio early — the free weights are the trial, the managed API is the commercial lane.
- Track Alibaba Cloud's AI-revenue growth (triple-digit for eight quarters) as the real scoreboard for whether open-weight giveaways convert into cloud revenue.