Introduction
Foundation models, pre-trained neural networks with capacity for transfer learning across multiple tasks, have become central to contemporary AI research and application. These models, characterised by their large parameter counts and broad generalisation capacities, underpin natural language processing, vision understanding and multimodal reasoning tasks. In the West, organisations such as OpenAI, Google DeepMind, Meta Platforms and Anthropic have driven much foundational research and deployment into consumer and enterprise services. However, Chinese tech firms are key participants in the global competitive ecology of foundational AI technology, shaped by state policy, industrial strategy and market forces.
Alibaba Group, a multinational technology conglomerate founded in 1999 by Jack Ma and associates, is best known internationally for its e-commerce platforms such as Taobao, Tmall and Alibaba.com. Over the past decade, however, AI and cloud computing have become central pillars in the company’s strategy for long-term technological leadership. Through concerted investments in AI research units such as the DAMO Academy and infrastructural commitments via Alibaba Cloud Intelligence, the company has developed a broad portfolio of AI capabilities. Among these, the Qwen family of large language and multimodal models stands out as a core expression of Alibaba’s vision for foundational AI systems that serve both industrial and research communities.
Early AI Development and Institutional Foundations
Alibaba’s journey into artificial intelligence did not begin with large language models but was rooted in earlier investments in data science, recommendation systems and computational infrastructure. As part of its e-commerce operations, Alibaba developed sophisticated machine learning systems to power search, recommendation, pricing and logistics optimisation. Over time, these operational AI systems laid the groundwork for more ambitious foundational research.
In 2016, Alibaba established its Artificial Intelligence Laboratory and in 2017 it launched the DAMO Academy (Discovery, Adventure, Momentum, Outlook), a global research initiative focused on exploratory science and engineering challenges, including AI and machine learning research. These organisational structures provided platforms for Alibaba to cultivate expertise and coordinate long-term research strategies beyond immediate commercial imperatives, signalling a deeper institutional commitment to AI as a strategic frontier.
Parallel to these research units, Alibaba Cloud (officially Alibaba Cloud Intelligence) expanded its infrastructure capabilities to support AI workloads with scalable distributed computing, storage and data processing services. Such infrastructure would later become essential for training and deploying large neural networks, underpinning the company’s progress into foundation models.
The Emergence of Qwen
Alibaba entered the field of foundation models in 2023 with the development of a large language model initially named Tongyi Qianwen (通义千问), often referred to by its project name Qwen. The release of a beta version occurred in April 2023 and after regulatory clearance in China, it was made publicly available in September 2023. This first generation marked Alibaba’s transformation from internal AI use cases towards generative AI with broad applicability. The architecture drew inspiration from open research such as the transformer-based architectures popularised in earlier foundation models and functioned as a robust pre-trained language model designed to handle both Chinese and English inputs.
Alibaba’s early Qwen models, such as Qwen-7B, Qwen-14B and larger variants like Qwen-72B and Qwen-1.8B, were designed to provide a spectrum of model sizes appropriate for different deployment scenarios, from cloud-hosted high-capacity inference to lighter models suitable for research experimentation and edge use. Many of these early models supported long context windows, enabling advanced handling of extended text segments beyond traditional limits and provided core natural language generation capabilities including summarisation, question-answering and conversational interfaces.
Open-Source Strategy and Ecosystem Building
Alibaba’s strategy with Qwen models has combined commercial cloud services with an open-source ethos that invited global adoption and derivative model development. In 2023, Alibaba open-sourced certain models, such as the 7 billion parameter Qwen-7B and Qwen-7B-Chat variants, making them available on platforms like ModelScope and Hugging Face. This approach enabled researchers and developers to experiment with foundation models without having to invest significantly in their own training processes, thereby facilitating broader ecosystem engagement.
However, while many models were described as open source, issues remain around the completeness of this openness: not all aspects of the training code and dataset documentation were released in ways that satisfy strict definitions of open-source AI model criteria, such as those outlined by the Linux Foundation’s Model Openness Framework. This nuance highlights tensions between corporate strategy and community expectations in open AI research practices.
Qwen2 and Multilingual Expansion
In June 2024, Alibaba introduced the Qwen2 series, reflecting iterative advancements over the initial Qwen generation. Qwen2 models included both dense and Mixture-of-Experts (MoE) variants that enabled conditional computational pathways, designed to improve performance and efficiency on reasoning, instruction following and multilingual comprehension tasks. These models incorporated improvements in training data diversity and architecture, making them competitive against prominent closed and open foundation models globally.
The Qwen2 models continued to expand support for multilingual capabilities, extending performance across a wider range of languages while enhancing contextual understanding and generation. In some configurations, these models also supported multimodal features when paired with appropriate vision-language components, demonstrating early integration of multimodal competencies within Alibaba’s foundation model ambitions.
Reasoning Models and QwQ-32B
In March 2025, Alibaba unveiled the QwQ-32B Preview model, a reasoning-oriented architecture described as offering enhanced problem-solving, logical inference and programmable capabilities. Positioned to challenge competitive models such as DeepSeek’s R1, QwQ-32B achieved performance comparable to leading systems in a range of scientific and coding benchmarks, emphasising Alibaba’s commitment to advancing reasoning quality in foundation models. The model’s release coincided with heightened interest from investors and regulatory stakeholders, reflecting the broader economic significance of AI innovation within China.
Qwen2.5 and Multimodal Development
Building on these advancements, the Qwen2.5 family, released in late 2024 and early 2025, represented a strategic deepening of Alibaba’s foundation model capabilities. These models introduced enhanced multimodal support, enabling not only text but also vision and audio integration and offered configurations such as Qwen2.5-VL and Qwen2.5-Omni that accepted diverse data types. Certain variants were made available under permissive licences (e.g. Apache 2.0), expanding integration possibilities for research and application.
Qwen3 and Next-Generation Scale
In April 2025, Alibaba released the Qwen3 family, representing a major generational leap in model capability and accessibility. The Qwen3 series comprises dense and MoE architectures with parameter scales spanning from modest to very large (e.g. 30 billion and 235 billion active parameters) and was trained on an exceptionally large multilingual corpus of around 36 trillion tokens across 119 languages and dialects. This scale and breadth underscore Alibaba’s ambition to build foundation models that are not only powerful but inclusive and globally relevant.
A key innovation in Qwen3 was the introduction of “hybrid reasoning” or dual-mode processing, enabling models to operate in a “thinking” mode for complex, multi-step tasks and a “non-thinking” mode for rapid, general-purpose responses. This duality reflects evolving research trends that foreground efficiency and dynamic reasoning within large neural models, positioning Qwen3 as a competitive alternative to contemporaneous systems from global competitors.
In September 2025, Alibaba unveiled Qwen3-Max, its most powerful AI model to date, containing over 1 trillion parameters and exhibiting significant performance on benchmarks for coding, reasoning and agent-like behaviour. The model’s capabilities in autonomous agent activity highlight Alibaba’s ambitions beyond static text generation toward more interactive and goal-oriented AI systems. These continued advancements are backed by a significant long-term infrastructure investment of around 380 billion yuan in cloud and AI development, illustrating the company’s recognition of foundational AI’s strategic importance in its broader portfolio.
Alibaba Cloud Integration and Developer Platforms
A distinctive aspect of Alibaba’s approach to foundation models is their tight integration with Alibaba Cloud’s infrastructure and service platforms. Tools such as Model Studio and the Model Context Protocol enable developers and enterprises to build, fine-tune and deploy AI applications using Qwen models without the need to train models from scratch. These platforms streamline AI adoption across sectors ranging from automotive and consumer electronics to healthcare and robotics, demonstrating how foundation models can enhance domain-specific processes and productivity.
The emphasis on cloud-native deployment is consistent with Alibaba’s technology strategy: AI capabilities are not isolated research curiosities but integral components of scalable, reliable infrastructure offerings for business customers. This strategy has helped spur adoption among hundreds of thousands of enterprises, illustrating AI technology’s role in broader economic digital transformation.
Multimodal and Industry Applications
Alibaba’s foundation model ecosystem supports not only text-centric automation but also multimodal applications. For example, vision-language models like Qwen-VL and other variants enable complex image understanding and querying tasks, extending generative AI into new interactive domains. Organisations in automotive manufacturing, smart device ecosystems, robotics and consumer services are deploying these capabilities to create enhanced user experiences and intelligent systems that operate across sensory modalities.
This breadth of application underscores how foundational AI models can catalyse innovation across industries when embedded within cloud, data and edge computing ecosystems, reflecting a shift towards pervasive AI integration in economic and societal functions.
Derivative Ecosystems and Open Adoption
Alibaba’s open-source strategy with Qwen and related models has catalysed widespread derivative activity. According to company reports, more than 300 AI models built on the Qwen and visual generation model frameworks have been open-sourced, collectively downloaded over 600 million times and spawning large derivative ecosystems. This scale of adoption signals that Alibaba’s foundation models have become one of the most widely used open-source AI series globally and that its approach to disseminating models has created substantial network effects.
From the perspective of technology governance, this open approach contrasts with more restrictive licensing seen in some proprietary systems and aligns with broader debates over democratising access to foundational AI resources. However, the debate around what constitutes sufficient openness, particularly regarding the disclosure of training data sources and complete code, remains unresolved in academic and industry circles.
Regulatory Context and Domestic Competition
Alibaba operates within a unique regulatory environment in China, where generative AI models must adhere to evolving government guidelines. Models intended for public deployment often require regulatory vetting and no foreign-developed models have yet been approved under certain Chinese guidelines as of late 2025. This regulatory context shapes the competitive landscape, favouring domestically developed models like Qwen, DeepSeek’s R1 and Tencent’s Hunyuan and generating distinct market dynamics compared to Western jurisdictions.
Moreover, competition among Chinese tech giants has intensified, particularly as startups such as DeepSeek achieve notable performance and cost efficiencies. Alibaba’s accelerated release cadence for Qwen developments reflects this competitive pressure, driving innovation even as geopolitical and chip export restrictions complicate the procurement of advanced hardware.
Cloud Computing, Enterprise Strategy and Socio-Technical Change
Alibaba’s foundational AI work is tightly coupled with its cloud computing strategy, a business unit that has grown into a core revenue and innovation engine. By providing powerful AI models through cloud platforms, the company reinforces its competitive position relative to international cloud providers and creates new revenue streams rooted in AI-enabled services.
This integration underscores a broader socio-technical transition where AI infrastructure, cloud ecosystems, enterprise adoption and digital transformation intersect. Foundation models are not merely research artefacts but infrastructural tools that reshape economic value chains, labour processes and organisational capabilities.
Conclusion
Alibaba’s evolution in artificial intelligence foundation models, beginning with Tongyi Qianwen in 2023 and progressing through iterative advances in Qwen2, QwQ reasoning models, Qwen2.5 multimodal systems and the next-generation Qwen3 and Qwen3-Max architectures, embodies a deliberate strategy that blends technological ambition with commercial and infrastructural deployment. The company’s approach combines open-source dissemination with cloud integration, enterprise adoption with broad ecosystem engagement and continuous research innovation with strategic industrial positioning.
Alibaba’s work illustrates how foundation models have become central not only to cutting-edge AI research but also to broader economic and societal transformations. By embedding these models within scalable cloud platforms and encouraging derivative innovation, Alibaba has contributed to an AI ecosystem that extends beyond proprietary systems into open and collaborative infrastructure, a significant trend in global AI development.
As Alibaba continues to advance foundational AI capabilities within a competitive and regulated environment, its trajectory offers rich material for further research on the economic, policy and ethical dimensions of large-scale AI model deployment. Understanding this trajectory enhances our comprehension not only of Alibaba’s technological strategies but also of how AI is reshaping global industrial and innovation landscapes.