Alibaba Cloud and Qwen

Alibaba Cloud is the cloud computing division of Alibaba Group and the largest provider of cloud services in China. Since 2023 it has also been the publisher of Qwen, a family of language and multimodal models that has become the most widely used Chinese open-weight family and one of the most widely derived in the world 1.

Qwen — Tongyi Qianwen in Chinese — began as an internal project and became a product line that spans models small enough to run on a phone and large enough to compete with the best American systems. Most of the family is released under the Apache 2.0 licence, which permits commercial use and modification, and the rest is sold through Alibaba Cloud's Model Studio platform 2.

The combination is unusual. Alibaba is a public company with an e-commerce business, a cloud business and a research organisation, and Qwen serves all three: it improves Alibaba's own products, it attracts developers to the cloud, and it gives the company a voice in the global argument about open models 3.

Alibaba Cloud, known in Chinese as Aliyun, was founded in 2009 to provide computing and storage for Alibaba's own commerce platforms. It grew into a public cloud with regions across Asia, Europe and the Middle East, and it holds roughly a third of the Chinese market, ahead of Huawei Cloud, Tencent Cloud and state telecom providers 1.

The business is strategically important to Alibaba because cloud revenue is recurring and because the company's commerce operations depend on it. It is also capital-intensive: data centres, servers and networking consume a large share of Alibaba's capital spending, and the AI build-out has increased that share 1.

Unlike its American counterparts, Alibaba Cloud has had to build much of its stack under export controls that limit access to the most advanced NVIDIA chips. It designs its own servers, has developed custom silicon through its PingTouge subsidiary, and has increasingly relied on domestic accelerators for inference 3.

Its international business is smaller than its domestic one but growing, particularly in Southeast Asia and the Middle East. Qwen is part of that pitch: a customer who builds on Qwen models through Alibaba Cloud is a customer of Alibaba's regions, storage and networking 4.

Qwen was announced in April 2023, in the same period as Baidu's ERNIE and a wave of Chinese chat products. The first open release, in August 2023, was a 7-billion-parameter model; the technical report that followed described base and chat models from 1.8 to 14 billion parameters, trained on multilingual data with a particular emphasis on Chinese 5.

Successive generations widened the range. Qwen1.5, in February 2024, added sizes from 0.5 to 72 billion parameters and a mixture-of-experts variant. Qwen2, in June 2024, improved multilingual performance and extended context. Qwen2.5, in September 2024, added specialised models for mathematics and coding, and QwQ-32B-Preview, in November, was the team's first reasoning model 6.

Qwen3, released in April 2025, was the most substantial redesign. It included two mixture-of-experts models — a 235-billion-parameter model with 22 billion active parameters and a 30-billion-parameter model with 3 billion active — and six dense models from 0.6 to 32 billion parameters, all under Apache 2.0. The models were trained on about 36 trillion tokens, support 119 languages, and can switch between a reasoning mode and a fast response mode 2.

Later releases extended the family into coding and into larger API-only models, including Qwen3-Coder and Qwen3-Max. The pattern is consistent: open weights for the sizes that developers can run, closed or API-only access for the largest and most capable systems 6.

Most Qwen models are released under Apache 2.0, one of the most permissive software licences. That allows companies to use them commercially, modify them and redistribute them without paying Alibaba or disclosing their changes. The Qwen3 release, which placed every size under Apache 2.0, was notable because several competitors had moved in the opposite direction 2.

The exceptions matter. Some earlier models, including certain 72-billion-parameter variants, were released under a separate Qwen licence that imposed conditions on redistribution, and the largest current models are available only through the API. The family is therefore open-weight rather than uniformly open-source, and the licence varies by model 6.

What is not released is as important as what is. Training data, data pipelines and the full training code are not published, so the models can be inspected and run but not reproduced. That is the norm for open-weight releases, and it is the main reason the Open Source Initiative's definition does not cover most of them 3.

The practical effect has been a large derivative ecosystem. Qwen models have been fine-tuned for dozens of languages, adapted to domain tasks, and used as the base for other companies' flagship models, particularly in Southeast Asia and the Middle East. On Hugging Face, Qwen-derived models are among the most numerous of any family 3.

Model Studio, known in Chinese as Bailian, is Alibaba Cloud's platform for serving Qwen models. It offers an API with per-token pricing, tools for fine-tuning and evaluation, and deployment options that range from shared endpoints to dedicated capacity. It is the commercial counterpart to the open weights 4.

The pricing is aggressive. Alibaba has cut Qwen API prices repeatedly, in some cases by an order of magnitude, in response to competition from DeepSeek, ByteDance's Doubao and others. The strategy is to make the API cheap enough that developers do not bother to self-host, while keeping the largest models available only through the platform 4.

For Alibaba, the return is indirect. Qwen attracts developers to Alibaba Cloud, where they buy compute, storage and databases; it improves Alibaba's own applications, including its e-commerce search and customer service; and it gives the company a credible claim to technical leadership in a market where Baidu, Tencent, ByteDance and Huawei are all competing 1.

The consumer side is smaller. The Tongyi application gives users access to Qwen models, but it competes with Doubao, Kimi and DeepSeek's own app in a market where consumer AI products are largely free and heavily subsidised 3.

Alibaba Cloud is a unit of Alibaba Group, which is listed in New York and Hong Kong and is one of China's largest companies by revenue. The cloud and AI businesses are reported within the group's accounts, and Alibaba has said that the cloud unit is central to its future growth as commerce matures 3.

In 2025 Alibaba committed at least 380 billion yuan, roughly 53 billion dollars, to cloud and AI infrastructure over three years, one of the largest such commitments by any company outside the United States. The money is going into data centres, accelerators and networking, and it is constrained by the availability of advanced chips under export controls 1.

The Qwen team sits inside the cloud group rather than in a separate laboratory. That gives it access to Alibaba's infrastructure and customers, and it subjects it to the priorities of a business unit with quarterly targets. The tension between research ambition and commercial delivery has surfaced repeatedly in the team's history 7.

Alibaba's ownership structure has also shaped the AI strategy. The company's dual-class share structure and its partnerships give management control, and its leadership has treated AI as a strategic priority since 2023, when it reorganised into six business groups and placed cloud and AI at the centre 3.

The Qwen team is part of Alibaba Cloud's research organisation and has been led by Lin Junyang, a researcher who led the team from its first open releases. The team publishes technical reports alongside releases, maintains the Qwen blog and repository, and engages directly with developers on social platforms 6.

Its engineering culture is distinct from that of Alibaba's product teams. The team releases weights, runs public benchmarks and responds to community requests, which has earned it a following among developers that few Chinese corporate research groups have. That following is an asset for Alibaba's cloud business and a source of independence for the team 7.

In March 2026 Lin announced that he was leaving the Qwen project. The departure was reported to have prompted an emergency question-and-answer session with Alibaba executives, and former colleagues said the team had long lacked computing and infrastructure support relative to its ambitions. Alibaba did not immediately name a successor 7.

The episode illustrated a structural issue for Chinese AI research. The best teams are small, mobile and expensive to retain, and they sit inside large companies whose priorities can change faster than a model roadmap. Alibaba's challenge is to keep the team that made Qwen distinctive while integrating its work into the cloud business 7.

The most documented criticism concerns political content. A 2026 study of Southeast Asian models built on Qwen found that they refused to answer certain politically sensitive questions and, in other cases, produced answers aligned with Chinese government narratives. The authors used a published list of sensitive questions developed by Chinese institutions and tested whether the models would answer them 8.

The behaviour is consistent with Chinese regulation. Generative AI services available to the public in China are required to uphold official positions on sensitive topics, and models trained in the country are shaped by those requirements. Alibaba has not published a detailed account of how its models are filtered, and the research suggests the filtering is not limited to explicit refusals but includes framing and omission 8.

The licensing and openness claims have also drawn criticism. Calling Qwen open source while withholding training data and releasing the largest models only through an API is a position that resembles that of Meta and Mistral; the Open Source Initiative's definition would not accept it. Alibaba describes the models as open and emphasises the permissive licences on the weights it does publish 2.

A third line of criticism concerns concentration. Qwen's dominance as a base model means that a single company's choices about training data and content policy propagate into hundreds of derivative models, including ones deployed in countries with different speech laws. That is a form of influence that no regulator currently oversees 3.

How open Qwen really is depends on which model is meant. The Qwen3 releases are Apache 2.0 and genuinely permissive; earlier and larger models are not, and the flagship systems are available only as a service. Both facts are in the technical reports, and the disagreement is about which should define the family 2.

Benchmark claims follow the industry pattern. Qwen3 was presented as competitive with DeepSeek-R1, OpenAI's o1 and o3-mini, Grok 3 and Gemini 2.5 Pro on mathematics, coding and general tasks, using the team's own evaluations. Independent testing has generally found the models strong for their size, with the usual caveats about contamination and prompt sensitivity 2.

A third disagreement concerns the March 2026 departure. It is either a serious loss that signals internal strain, or a normal executive move in a competitive talent market. Alibaba's public response emphasised continuity, and the models continued to ship, but the reports of resource constraints within the team were not denied 7.

Finally, there is the question of what Alibaba gains. Open weights do not generate revenue directly, and the API business is priced low. The company's bet is that Qwen's adoption pulls customers into its cloud. Whether that bet pays for 380 billion yuan of infrastructure is not yet known 1.

Qwen is China's most widely used open-weight family and one of the two or three most important sources of open models in the world, alongside Meta's Llama and Mistral's releases. Its Apache 2.0 licences made it the default choice for developers who needed a capable model without an American vendor 3.

Its position rests on Alibaba's infrastructure and on a team that has been unusually willing to publish. Both are advantages that competitors can copy, and the American export controls limit the compute available to train the next generation. The company's response has been to optimise aggressively for the hardware it can obtain, a constraint that has shaped Chinese model design more broadly 1.

For the encyclopedia, Alibaba Cloud and Qwen represent the open-weight strategy pursued at platform scale. The models are a loss leader for the cloud, a statement of technical capability, and a form of influence over how Chinese-language AI is built. The commercial and political returns are real, and they are not the ones a Western open-source project would recognise 4.