Alibaba AI
Last reviewed
May 1, 2026
Sources
35 citations
Review status
Source-backed
Revision
v1 · 3,998 words
Improve this article
Add missing citations, update stale details, or suggest a clearer explanation.
Suggest edit
CtrlK
Last reviewed
May 1, 2026
Sources
35 citations
Review status
Source-backed
Revision
v1 · 3,998 words
Add missing citations, update stale details, or suggest a clearer explanation.
Alibaba AI refers to the artificial-intelligence work of Alibaba Group Holding Limited, a Chinese multinational technology conglomerate headquartered in Hangzhou. Alibaba's AI activities are organised primarily around three vehicles: DAMO Academy, the company's central research arm founded in October 2017; Alibaba Cloud, the cloud-services subsidiary that began operations in September 2009 as Aliyun and now ranks fourth among global infrastructure-as-a-service providers; and the Qwen family of large language models, originally introduced in April 2023 under the consumer brand Tongyi Qianwen [1][2][3].
Since late 2023 Alibaba has positioned AI as the centre of its corporate strategy. CEO Eddie Wu, who took office on 10 September 2023 alongside chairman Joseph Tsai, has repeatedly described the company as pursuing an "AI-driven, user-first" agenda, and on 24 February 2025 the firm committed at least 380 billion yuan (about US$53 billion) over three years to AI and cloud infrastructure, the largest single private computing project ever announced in China [4][5][6]. The Qwen models, released mostly under permissive open-weight licences, have made Alibaba arguably the most-downloaded open-weight LLM provider in the world, alongside Meta's Llama family, with the Qwen3 series alone training on roughly 36 trillion tokens spanning 119 languages [7][8].
Alibaba Group was founded in 1999 in Hangzhou by Jack Ma and seventeen co-founders as a B2B marketplace connecting Chinese exporters with overseas buyers. The group listed on the NYSE in September 2014 under the ticker BABA in what was then the largest IPO in history, and added a secondary listing on the Hong Kong Stock Exchange (9988) in November 2019 [1]. Today its consumer e-commerce platforms (Taobao, Tmall, AliExpress, Lazada), its cloud and AI businesses (Alibaba Cloud, DAMO Academy, ModelScope), its logistics arm Cainiao, and its digital-media holdings sit under one roof, while the financial-services affiliate Ant Group operates separately under common-control links.
The leadership reshuffle of 2023 was the proximate trigger for Alibaba's current AI focus. On 20 June 2023 the company announced that Daniel Zhang would step down as chairman and CEO; Joseph Tsai (Cai Chongxin), an Alibaba co-founder and longtime executive vice chairman, would become chairman; and Eddie Wu Yongming, another co-founder who had served as the group's first technology director, would become CEO. Both transitions took effect on 10 September 2023 [4]. Wu also took direct charge of Alibaba Cloud Intelligence Group in September 2023 and personally oversaw Taobao and Tmall Group from December 2023, running Alibaba's two largest engines as part of one AI-led story [9].
Key AI executives include:
| Name | Role | Notes |
|---|---|---|
| Eddie Wu (Wu Yongming) | CEO | Co-founder, former CTO; assumed CEO role 10 Sep 2023; champion of "AI-driven, user-first" strategy [4] |
| Joseph Tsai (Cai Chongxin) | Chairman | Co-founder; assumed chairman role 10 Sep 2023; previously executive vice chairman [4] |
| Jeff Zhang (Zhang Jianfeng) | DAMO Academy president | Group CTO at the founding of DAMO Academy in 2017 [10] |
| Junyang Lin | Qwen technical lead (until March 2026) | Lead author on Qwen, Qwen2, Qwen2.5 and Qwen3 technical reports; departed 3 March 2026 [11][12] |
| Cheng Yongchen | Tongyi product lead | Heads the Tongyi consumer-facing AI products inside Alibaba Cloud Intelligence |
The Academy for Discovery, Adventure, Momentum and Outlook, universally known as DAMO Academy, was launched on 11 October 2017 at Alibaba's Computing Conference in Hangzhou. Then-CTO Jeff Zhang announced a commitment of more than US$15 billion in research-and-development spending over three years, and said the academy would build seven labs in Beijing, Hangzhou, San Mateo, Bellevue, Moscow, Tel Aviv and Singapore, recruiting roughly 100 researchers in the initial wave [10][13]. The advisory board was deliberately international, drawing professors from Princeton, Harvard, MIT, the University of Washington, Columbia, Beijing Institute of Technology, Peking University and Zhejiang University, with collaboration ties to UC Berkeley's RISE Lab [10].
DAMO's stated remit covers data intelligence, the Internet of Things, financial intelligence, quantum computing and human-machine interaction, with deeper specialisations in machine learning, natural language processing, computer vision, network security and autonomous driving [10][14]. Notable research outputs over the years include AliceMind (a family of pre-trained language models open-sourced in 2021), the OFA "One-For-All" 6-billion-parameter cross-modal model, and the M6 series of multimodal multitask transformers. M6 itself went through three jumps in 2021: a 100-billion-parameter version in March, a 1-trillion-parameter version in June, and a 10-trillion-parameter sparse version announced in November, which Alibaba claimed it trained on 512 GPUs while consuming roughly 1 percent of the energy of a comparable GPT-3-class run [15][16]. The PLUG and ChatPLUG dialogue models followed in 2022 and early 2023 and laid groundwork for the eventual Tongyi Qianwen launch.
Not every part of the original blueprint survived. In November 2023 DAMO Academy closed its quantum-computing laboratory, donating its research equipment to Zhejiang University and saying the academy would refocus on artificial intelligence [17]. Several of the international labs have been quietly restructured during the period of US-China technology tensions, although the Sunnyvale and Seattle offices remained active as of late 2025.
Tongyi Qianwen, which translates roughly as "truth seeking by asking a thousand questions," is the consumer brand under which Alibaba launched its first generative-AI chatbot. The product was unveiled on 11 April 2023 at a press event in Beijing by then-CEO Daniel Zhang, with Alibaba positioning it as a Chinese alternative to ChatGPT capable of working in both Chinese and English [18][19]. Initial integration targets were Alibaba's workplace tool DingTalk and the Tmall Genie smart-home line; public access opened in September 2023 after Chinese regulators completed their approval process [7][19].
The English-language model brand Qwen (a contraction of Qianwen) is now used for the open-weight series. Open releases began in August 2023 with Qwen-7B and the multimodal Qwen-VL, expanded across the rest of 2023 with Qwen-14B (September), Qwen-Audio (November) and Qwen-72B plus Qwen-1.8B (December), and have continued at a rapid cadence ever since [7][20].
| Release | Date | Sizes | License | Notes |
|---|---|---|---|---|
| Tongyi Qianwen beta | 11 Apr 2023 | n/a (closed) | Proprietary | First Alibaba LLM chatbot, unveiled in Beijing [18] |
| Qwen-7B | Aug 2023 | 7B | Tongyi Qianwen Research | First open-weight Qwen model on Hugging Face and ModelScope [7] |
| Qwen-VL | Aug 2023 | 7B-class | Tongyi Qianwen Research | Vision-language model based on a ViT plus Qwen-7B [7] |
| Qwen-14B | 25 Sep 2023 | 14B | Tongyi Qianwen Research | Released alongside chat variant on ModelScope and Hugging Face |
| Qwen-Audio | 30 Nov 2023 | 7B-class | Tongyi Qianwen Research | Speech, music and natural sound to text; covers 30+ tasks in 8 languages [21] |
| Qwen-72B and Qwen-1.8B | Dec 2023 | 72B, 1.8B | Tongyi Qianwen Research | Largest Chinese open-weight LLM at the time [7] |
| Qwen1.5 | 4 Feb 2024 | 0.5B, 1.8B, 4B, 7B, 14B, 32B, 72B (MoE added later) | Mostly Tongyi Qianwen | Switched architecture closer to Llama family [22] |
| Qwen1.5-110B | 25 Apr 2024 | 110B | Tongyi Qianwen | First Chinese open-weight model above 100B parameters; competitive with Llama-3-70B base [23] |
| Qwen2 | 6 Jun 2024 | 0.5B, 1.5B, 7B, 57B-A14B, 72B | 72B Tongyi Qianwen, others Apache 2.0 | Native 128K context for 7B and 72B Instruct; first MoE in main line [24] |
| Qwen2-Audio | Aug 2024 | 7B-class | Apache 2.0 | Successor to Qwen-Audio with broader instruction tuning |
| Qwen2-VL | Aug 2024 | 2B, 7B, 72B | Apache 2.0 (2B/7B) | Improved vision-language and video understanding |
| Qwen2.5 | 19 Sep 2024 | 0.5B, 1.5B, 3B, 7B, 14B, 32B, 72B | Apache 2.0 except 3B and 72B | Trained on 18T tokens; 29+ languages; 128K context [25] |
| Qwen2.5-Coder | Nov 2024 | 1.5B, 7B, 32B | Apache 2.0 | Trained on 5.5T code tokens; flagship for open coding models |
| Qwen2.5-Math | Sep 2024 | 1.5B, 7B, 72B | Apache 2.0 (1.5B, 7B) | Math-specific reasoning, supports CoT, PoT and tool use |
| QwQ-32B-Preview | 27 Nov 2024 | 32.5B | Apache 2.0 | Open challenger to OpenAI's o1-preview and o1-mini in reasoning [26] |
| Qwen2.5-VL | Jan 2025 | 3B, 7B, 32B, 72B | Apache 2.0 (most) | Vision-language follow-up; runs on consumer hardware at smaller sizes |
| Qwen2.5-Omni | 26 Mar 2025 | 7B | Apache 2.0 | End-to-end multimodal speech, vision and text in one model |
| QwQ-32B | Mar 2025 | 32B | Apache 2.0 | Full release of the QwQ reasoning line, reinforcement-learning trained over Qwen2.5-32B |
| Qwen3 | 28-29 Apr 2025 | Dense 0.6B, 1.7B, 4B, 8B, 14B, 32B; MoE 30B-A3B, 235B-A22B | Apache 2.0 | Trained on 36T tokens, 119 languages; integrates a hybrid "thinking" and "non-thinking" mode [8][27] |
| Qwen3.5 | Feb 2026 | Dense and MoE variants | Apache 2.0 (open) | Open-weight base; Qwen3.5-Plus is proprietary [7] |
| Qwen3.6 | Apr 2026 | Dense 35B-A3B and Plus | Apache 2.0 (35B-A3B); Qwen3.6-Plus proprietary | Latest generation as of writing [7] |
The early Qwen weights shipped under the company's own "Tongyi Qianwen Research" agreement, which was free for research and required a separate application for commercial use beyond a 100-million monthly-active-user threshold. From Qwen2 onwards the bulk of the family has shipped under Apache 2.0, with only the largest 72B and a few specialised variants (such as Qwen2.5-3B) keeping a more restrictive licence [7][24]. By late 2025 Alibaba had released over 100 open-weight Qwen variants and the family had accumulated more than 600 million downloads on Hugging Face, with over 170,000 derivative models built on top of Qwen, surpassing Meta's Llama by both metrics [11][28].
Qwen models are designed around a Transformer decoder architecture. The original Qwen-7B used a standard rotary-position-embedding setup; from Qwen1.5 onward the team aligned more closely with the Llama-style stack, including SwiGLU activations, grouped-query attention in larger sizes, and Mixture-of-Experts in Qwen2-57B-A14B, Qwen3-30B-A3B and Qwen3-235B-A22B [22][24][8].
Key capabilities span:
Flagship Qwen3-235B-A22B has activated 22 billion of its 235 billion parameters per token and reaches 85.7 on AIME'24, 81.5 on AIME'25, 70.7 on LiveCodeBench v5 and 2,056 on Codeforces, scores that the Qwen team reports as competitive with DeepSeek-R1 and OpenAI o1-mini [8][27].
Alibaba Cloud, originally Aliyun, opened for business in September 2009 and grew into China's largest cloud provider. By 2025 it remained the fourth-largest infrastructure-as-a-service vendor globally, behind Amazon Web Services, Microsoft Azure and Google Cloud Platform, with worldwide IaaS share rising to 7.7 percent (up from 7.2 percent in 2024) and Asia-Pacific share at 22.5 percent according to Gartner's April 2026 figures [29]. The company operates 78 data centres across the Asia-Pacific region and ranks first by revenue in mainland China and Hong Kong, second in Indonesia and Malaysia, and third in Singapore [29].
The firm's AI products sit on two main platforms. Platform for AI (PAI) is the long-running suite for end-to-end machine-learning workflows, including PAI-DSW (data-science workshop), PAI-Studio for visual workflow design, and PAI-EAS (Elastic AI Serving) for scalable model deployment [30]. Model Studio is Alibaba Cloud's generative-AI platform launched in 2023 and expanded internationally in 2024, exposing Qwen, Qwen-VL and the visual-generation model Tongyi Wanxiang via APIs comparable to those of OpenAI or Anthropic [30][31].
On the open-source side, ModelScope opened on 3 November 2022 at the Apsara Conference as a Hugging-Face-style "Model-as-a-Service" community, initially with about 300 ready-to-deploy models drawn from DAMO Academy [32]. By 2025 it had grown into China's largest open-source AI community, hosting more than 70,000 models and serving more than 16 million developers across 36 countries [32].
At the chip layer, Alibaba's semiconductor subsidiary T-Head (formally known as Pingtouge) has been designing custom silicon since 2018. Its first AI chip, Hanguang 800, was unveiled at the 2019 Apsara Conference: a 12-nanometre, ~17-billion-transistor inference accelerator that hit 78,563 images per second on the ResNet-50 benchmark and 500 inferences-per-watt for power efficiency [33]. Hanguang has since been used for image search and city-management workloads inside Alibaba; in production AI training Alibaba relies on a mix of NVIDIA GPUs (within the limits of US export controls, including the H20 part shipped to China), Huawei Ascend accelerators and its own custom designs.
| Item | Detail |
|---|---|
| Announcement date | 24 February 2025 [5][6] |
| Headline figure | At least RMB 380 billion (about US$53 billion) [5][6] |
| Time horizon | Three years (2025-2028) |
| Coverage | Cloud and AI infrastructure, including data centres, GPUs and networking |
| Historical context | Exceeds Alibaba's combined cloud-infrastructure spending over the previous decade [5] |
| Comparison | Roughly half of the US$100 billion announced for the US Stargate project; comparable to capex plans of major US hyperscalers [5] |
| Subsequent revisions | In late September 2025 Alibaba indicated it would raise the total above the original US$50 billion figure [34] |
Eddie Wu, on the analyst call accompanying the announcement, said Alibaba would "aggressively invest" in the new build-out and tied the spending plan to a five-year revenue target of US$100 billion in cloud and AI [5][6].
Beyond the cloud and the Qwen models, Alibaba ships AI capabilities through the broader portfolio. Taobao and Tmall use machine-learning recommendation engines for ranking, search and personalised marketing, and added more than ten generative-AI features for merchants in 2024 along with a generative chat assistant called Wenwen for shoppers [9]. The workplace tool DingTalk integrates a Tongyi-powered assistant for meeting summaries, document drafting and code generation, and the low-code platform Modao surfaces AI-generated UI components.
The logistics arm Cainiao runs AI-driven warehouse and last-mile operations: a 30,000-square-metre automated facility uses around 1,000 autonomous mobile robots to handle 34,000 SKUs at peak picking rates of 600 pieces per station per hour and a reported 99.99 percent accuracy [35]. In 2026 Cainiao began rolling out a global network of robotic warehouses across Hong Kong, mainland China, the Netherlands, Spain, France, Germany and the United States, paired with an AI scheduling system for routing and collision avoidance [35]. The related but separately operated Ant Group runs Alipay's biometric-payment, fraud-detection and insurance-underwriting models. Inside Alibaba Cloud, the visual-generation model Tongyi Wanxiang covers text-to-image and text-to-video, and the speech-recognition stack (QASR) handles real-time transcription for contact-centre and live-streaming workloads.
Alibaba is one of several major AI players in China, and increasingly compares directly with American hyperscalers and frontier labs. The table below sketches the landscape as of early 2026.
| Company | Founded | Flagship LLM family | Open weights | Multilingual | Distinguishing strength | Cloud presence |
|---|---|---|---|---|---|---|
| Alibaba | 1999 | Qwen / Tongyi Qianwen | Yes (mostly Apache 2.0) | 119 languages (Qwen3) | Largest open-weight Chinese family, broad MoE catalogue | Alibaba Cloud, #4 global IaaS [29] |
| Tencent | 1998 | Hunyuan-Large, Hunyuan-DiT | Some open weights, mostly closed | Yes | Strong creative and gaming integration | Tencent Cloud |
| Baidu | 2000 | ERNIE 4.0, ERNIE 4.5, ERNIEbot | Limited open weights | Yes | Search and conversational AI heritage | Baidu AI Cloud |
| DeepSeek | 2023 | DeepSeek V3, DeepSeek R1 | Yes (open MIT) | Yes | High efficiency, reasoning models trained at low cost | None of its own |
| 01.AI (Lingyiwanwu) | 2023 | Yi family | Yes | Yes | Bilingual Chinese-English models | None |
| Moonshot AI | 2023 | Kimi K1, K2 | Mostly closed | Yes | Long-context reading and search | None |
| MiniMax | 2021 | MiniMax-Text-01 (MoE) | Recently more open | Yes | Long-context MoE models | None |
| Zhipu AI | 2019 | GLM family | Mixed | Yes | Academic origin (Tsinghua) | None |
| Huawei | 1987 | Pangu | Closed | Yes | Vertical industry models, Ascend chips | Huawei Cloud |
| ByteDance | 2012 | Doubao | Mostly closed | Yes | Consumer scale, video and recommendation | Volcano Engine |
| OpenAI | 2015 | GPT-4o, o1, o3 | Limited open weights (gpt-oss family) | Yes | Frontier closed models, ChatGPT scale | None directly |
| Anthropic | 2021 | Claude 3.5/4 series | Closed | Yes | Constitutional AI, safety focus | None directly |
| 1998 (DeepMind 2010) | Gemini | Some Gemma open | Yes | Multimodal frontier and TPU stack | Google Cloud | |
| Meta | 2004 | Llama 2/3/4 | Yes (custom permissive) | Yes | Largest Western open-weight family | None directly |
In 2025 the Qwen family overtook Meta's Llama in raw monthly downloads on Hugging Face and in the count of fine-tuned derivative models, making Alibaba the most cited single source of open-weight LLM capacity globally [11][28]. DeepSeek emerged in early 2025 as the most disruptive Chinese rival, with V3 and R1 showing that frontier-grade reasoning could be trained on roughly US$5.6 million in compute [28]. Qwen is the more prolific publisher of model variants and the more widely fine-tuned base; DeepSeek tends to set the high-water mark for individual benchmark scores at any given time.
Alibaba's AI work happens against the backdrop of the wider China-US technology contest. From October 2022 onward, US export controls have tightened access to NVIDIA's most advanced datacentre GPUs (A100, H100 and successors), forcing Chinese cloud providers to rely on de-rated parts such as the H20, on Huawei Ascend hardware, or on domestic designs such as Hanguang 800. The same export regime has pushed Alibaba toward more aggressive efficiency engineering, which is one reason its mid-sized models compete on benchmarks with much larger Western systems.
Domestic policy has cut both ways. The Chinese government's 2017 "Next-Generation Artificial Intelligence Development Plan," the 2024 "AI+" initiative, and the sandbox programmes for foundation-model approval have all helped the case for sustained investment. The post-2020 regulatory crackdown on platform companies, which forced the suspension of Ant Group's IPO in November 2020 and produced a record antitrust fine for Alibaba in April 2021, weighed heavily on the firm's stock and capital plans for several years. The 2023 leadership change and the early-2025 stock recovery (following both DeepSeek's R1 launch and the US$53 billion infrastructure announcement) marked the first sustained rebound of the post-crackdown era, with the AI narrative doing most of the work.
In February 2024 Eddie Wu sent a widely reported internal memo defining "AI-driven" and "user value" as Alibaba's two strategic anchors, and in the following twelve months the company pushed Qwen releases out at a near-monthly cadence. The Qwen2 family arrived on 6 June 2024 and Qwen2.5 on 19 September 2024, both with permissive licensing on most variants, while Tongyi Qianwen 2.0 and industry-specific Tongyi models were unveiled at the October 2023 Apsara Conference and refreshed annually [25][30].
The February 2025 infrastructure commitment, paired with QwQ-32B-Preview in late November 2024 and the Qwen3 launch on 29 April 2025, pushed BABA stock to four-year highs through the spring and again in September 2025 when Alibaba signalled it would raise total AI capex above the original US$50 billion figure [34]. By the time Qwen3 shipped, cumulative downloads across the Qwen family had crossed 600 million on Hugging Face and Alibaba was hosting over 70,000 models on ModelScope [11][32].
Qwen3.5 small models launched in February 2026, and on 3 March 2026 longtime Qwen technical lead Junyang Lin and two other senior researchers departed in a brief exit announced via Lin's X post: "me stepping down. bye my beloved qwen" [11][12]. Alibaba consolidated its open-source efforts under a new Alibaba Token Hub. Qwen3.6 followed in April 2026, with the open Qwen3.6-35B-A3B under Apache 2.0 and Qwen3.6-Plus offered through the Tongyi consumer chat and Alibaba Cloud APIs.