Alibaba Cloud Open Source Large Model: Promoting AI Ecological Construction and Multi scenario Applications With the emergence of a large number of open source models, their performance can now be comparable to the most advanced closed source models. At the same time, the open source model effectively reduces the cost of model training, allowing the most advanced models to be deployed at low cost and widely applied.
In the era of AI, Alibaba Group has completely transformed its corporate development strategy.
On February 24, 2025, Alibaba Group CEO Wu Yongming announced that Alibaba plans to invest over 380 billion yuan in global cloud services and AI hardware infrastructure construction over the next three years, far exceeding the company's total investment in this field over the past decade.
In the field of large models, Alibaba Cloud has achieved remarkable results. At present, over 200 different models have been open sourced, and the number of Qwen derived models has exceeded 100000 in the open source community at home and abroad, surpassing the Llama series models in the United States and firmly ranking first in the global open source big models.
The Alibaba Cloud Tongyi Qianwen team revealed to Business School magazine that with the help of open source, Alibaba Cloud is committed to building a powerful AI ecosystem, helping a large number of small and medium-sized enterprises and AI developers to use Tongyi Qianwen earlier and more conveniently, promoting the popularization and implementation of big model technology, and contributing to the construction of China's big model ecosystem. Only by building a solid foundation of large-scale models can infinite possibilities be created for future commercial applications.
Fully open source to promote the construction of AI ecosystem
Starting from August 2023, Alibaba Cloud has taken the lead in large-scale open source of large model series in China, becoming the first large-scale open source manufacturer in China, and has won the titles of "Global Open Source Champion" and "Domestic Model Champion" on authoritative lists such as Chatbot Arena and Sinan OpenCompass multiple times. In 2024, the Qwen2.5-1.5B open source model alone will account for 26.6% of the world's largest open source community Hugging Face model downloads, ranking first in the world.
Alibaba Cloud pointed out that open source can gather the wisdom and strength of global developers, and compared to independent research and development by a single company, it can more efficiently promote technological innovation and iteration. For example, after DeepSeek opened up R1, there was a global craze for replicating R1, which had an impact on OpenAI's o1 inference model and propelled global big model technology into a new stage of inference models.
In addition, open source is beneficial for building ecosystems. In open source mode, anyone can download and deploy AI models for free, which effectively promotes the popularization of new technologies. Four of DeepSeek's six distillation models are based on the Qianwen open-source model; The Li Feifei team also successfully trained a new model s1-32B based on the Qianwen open source model, which has performance comparable to top-notch inference models. Its performance on competitive mathematical problems is 27% higher than o1 proof.
Furthermore, open source accelerates the industrialization process of AI and promotes the landing of AI technology in various industries. For startups and small and medium-sized enterprises, directly using open source models can significantly reduce the technological and financial barriers to entering the AI field. By combining their own knowledge base or corpus, they can quickly develop vertical solutions suitable for various industries.
Of particular note is the QwQ-32B released by Alibaba Cloud on March 6, 2025, which is not only suitable for enterprises but also convenient for individual developers and ordinary users to use. It can also be deployed locally on consumer grade graphics cards, significantly reducing deployment and usage costs. Through large-scale reinforcement learning, the model has achieved a leap in mathematics, code, and general abilities, with overall performance comparable to DeepSeeker R1.
However, as the marginal benefits brought by the Scaling Law gradually decrease, large model vendors represented by OpenAI have approached the limits of existing technology and resource conditions in terms of data accumulation and computing power stacking. Therefore, exploring new optimization paths has become crucial. The industry is beginning to reduce costs through extreme computing power optimization and improve in the reinforcement learning and inference stages to train high-performance models in a more efficient way. The launch of QwQ-32B once again confirms the enormous potential of reinforcement learning in improving model performance.
Alibaba Cloud stated that with the continuous emergence of open-source models, its performance can now rival the most advanced closed source models. The open source model also reduces the cost of model training, allowing the most advanced models to be deployed and applied at low cost.
Interestingly, in response to competition from Chinese companies, international giants such as Google and OpenAI have also begun to adjust their strategies and attempt limited open source. For example, Google previously focused on the closed source model Gemini, but recently opened sourced Gemma 3; OpenAI, once jokingly known as CloseAI, has been influenced by the open source environment in China. Its CEO Sam Altman admitted that the closed source strategy was wrong and announced future open source plans.
'Cloud+AI' prepares for application explosion
At present, more and more enterprises tend to directly use the standard API services provided by large model vendors, especially for scenarios where the business is relatively fixed and does not require extensive customization. The cost of using API services is more economical as it only incurs expenses during actual calls, making it particularly suitable for scenarios with obvious peaks and valleys in demand, such as customer service systems. Except for peak hours during the day, the demand is lower at other times, and there is no need to develop a large model from scratch, thereby significantly reducing costs.
Alibaba Cloud Bailian Platform is an AI big model service platform specially designed by Alibaba for enterprises, helping enterprises and developers to build, train, and deploy AI applications more efficiently. At present, API calls based on this platform are the main way for Alibaba Cloud's large-scale model commercialization. The Alibaba Cloud Bailian platform provides a fast and convenient service access method, making it easy for users to utilize Alibaba Cloud's large model capabilities.
In addition to standard API calls, Alibaba Cloud also provides a complete set of AI customized solution services jointly developed, covering model fine-tuning, customized models, etc., to meet customers who are unfamiliar with model technology or have highly customized needs. For users with data security and privacy protection needs, Alibaba Cloud has launched a private deployment solution, such as integrating models into all-in-one machines and selling them directly to customers.
In promoting the commercialization of open source big models, Alibaba Cloud continues to explore more cost-effective ways to open up AI capabilities.
Since 2024, Alibaba Cloud has repeatedly lowered the prices of the Tongyi Qianwen series models, with the input price of the Qwen Long API dropping from 0.02 yuan/thousand tokens to 0.0005 yuan/thousand tokens, a decrease of 97%. Other models such as Qwen Turbo, Qwen Plus, Qwen Max, and visual understanding models have also seen significant price reductions, allowing more businesses and developers to acquire AI capabilities at lower costs.
It is worth noting that the continuous cost reduction is due to Alibaba Cloud's continuous progress in AI infrastructure, especially in improving computing efficiency, optimizing network performance, and accelerating storage speed. For example, Alibaba Cloud has launched the high-performance network architecture HPN7.0 that supports stable connections of over 100000 GPUs, the Panju AI server that supports up to 16 cards per machine, and the CPFS file storage with a data throughput of 20TB/s.
In addition, the Alibaba Cloud Bailian platform has integrated over 200 large model APIs, covering mainstream open source and closed source models both domestically and internationally. To reduce the cost of users using the Text Large Model API, Alibaba Cloud's Bailian platform has launched a new KV Cache billing model, which automatically caches context to avoid duplicate calculations and further reduces model call costs. It is suitable for scenarios such as long texts, code completion, multi round conversations, and specific text summaries.
Overall, AI open-source big models face two major pain points in the commercialization process: firstly, the lack of "killer applications" that can significantly improve user experience or efficiency, which means that the market is not fully prepared to embrace AI big models, and their application scenarios and potential need to be further explored and verified; Secondly, the B-end market is still in the initial exploration stage, and the management of enterprises has insufficient understanding of using open source big models to achieve business value. Compared with closed source big models, it is more difficult to persuade them to invest in AI solutions.
Alibaba has prepared for these challenges and future plans through an integrated full stack solution of "cloud+AI". For example, utilizing the product power of AI infrastructure to expand the B2B market. At the same time, through the revision and upgrading of Quark, we are exploring the potential of large-scale model applications in the C-end market, aiming to create phenomenal product samples and showcase the best AI experience entrance.
Multi scenario exploration of Tongyi Qianwen
With the continuous advancement of big model technology, it has demonstrated enormous application value and potential in multiple scenarios.
In January 2024, China FAW and Tongyi Qianwen jointly launched the first business intelligence (BI) application based on big models in the automotive industry - GPT-BI. FAW utilizes the Qianwen big model capability to achieve data intelligence, transforming a large amount of dormant data assets into valuable resources.
Specifically, GPT-BI helps FAW achieve automatic report generation of sales data, greatly improving work efficiency. Meng Xin, Vice President of FAW Hongqi Brand Operation Committee, stated that there are still more application scenarios for AI in practical work. For example, in the process of vehicle development, a large amount of data is involved, including size matching, process review, assembly strategy, etc. in automobile production, which originally relied on manual measurement. Nowadays, based on these data and requirements, Hongqi has generated an automatic evaluation model for Hongqi automobiles and introduced it into the entire vehicle production process, reducing the previously time-consuming steps of 20, 50, or even 80 days to seconds or days on the workbench.
Based on the exploration of real business scenarios such as GPT-BI, FAW also plans to utilize a large amount of high-quality data assets to create vertical large models for multiple fields such as research and development, manufacturing, and after-sales service on the "Alibaba Cloud Bailian" one-stop large model application development platform, so that the large models can fully play their role in production, sales, and other aspects.
On March 11, 2025, Manus platform announced a strategic partnership with Alibaba's Tongyi Qianwen team. Both parties will implement all functions of Manus on domestic models and computing platforms based on the Tongyi Qianwen series open source model, committed to creating more creative general intelligent agent products for Chinese users.
The industry generally believes that Manus' collaboration with Alibaba is not only about utilizing advanced large-scale models, but also about finding cost-effective cloud platforms to meet the huge computing power demand, especially in the Chinese market. The cooperation between the two parties will accelerate the industrialization process of AI intelligent agents, reduce operating costs and improve service performance through efficient and economical cloud computing solutions.
Not long ago, the globally renowned technology giant Apple reached a cooperation agreement with Alibaba, with Alibaba serving as an Apple supplier to develop artificial intelligence features for the Chinese version of the iPhone.
Analysts pointed out that Apple chose to cooperate with Alibaba, on the one hand, because of the computing power advantage of Alibaba Cloud platform, and on the other hand, because the Tongyi Qianwen model series has strong multimodal capabilities. For AI agents, in addition to powerful language models, they also need to have the ability to understand multimodal inputs such as visual and audio, which is one of the key considerations for companies such as Apple when choosing AI big model partners.
In terms of multimodal capability, the Tongyi Qianwen Qwen Vision Language model performs well. Taking the latest version of Qwen2.5-VL as an example, this model can not only handle text input and output, but also has the ability to handle visual and video input, including understanding videos up to 1 hour long.
Compared to the previous generation Qwen2-VL, this model has undergone multiple updates. For example, the enhanced Grounding ability improves the recognition ability of physical events and objects, enabling the model to better associate textual descriptions with elements in actual images or videos. The enhanced OCR (Optical Character Recognition) capability performs well in handling complex or handwritten text, long texts, without the need for additional tuning for specific types of text, and performs better in handling text with scattered angles.
In Qwen2.5-VL, the Tongyi Qianwen team also designed a more comprehensive document parsing format - QwenVL HTML format, which can accurately recognize the text in the document and extract the position information of document elements (such as images, tables, etc.), accurately restoring the layout of the document.
It is reported that in addition to strong performance, the model also has built-in agent capabilities, such as the ability to output documents in standard JS format, which is crucial for the practical application of intelligent agents. Only by standardizing the output format can the agent effectively call these results. Through these clever designs, not only has the functionality and practicality of the model been enhanced, but it has also provided strong support for the future development of the AI industry.
Starting from 2024, Alibaba Cloud will deploy large models directly to devices such as smartphones to enable local task execution, such as voice summarization generation. The cooperation between Alibaba and Apple adopts an end-to-end cloud architecture, with large-scale models in the cloud and smaller parameter models configured on the device side, combined with security measures. This is a key development direction for the future.
From the deployment mode perspective, there are two main ways to deploy large models on mobile phones: one is local deployment, which embeds the large model into the phone, or deploys smaller models (such as 0.5B to 7B parameter scale) to perform data sensitive tasks; The second is cloud integration, where smaller models run locally to process sensitive data, while complex tasks are submitted to larger models in the cloud through the network for processing.
The advantage of cloud based big models lies in their powerful performance, ability to handle complex tasks, multiple parameters, and good performance, but there are issues with latency and data transmission privacy. Local large models provide better data security and privacy protection, with all processing completed on the device. They are sufficient for basic tasks, but due to size limitations, their performance may not be as good as cloud models, and the experience may be slightly inferior.
Although local models can meet basic task requirements, cloud based models are still indispensable for customized needs with higher levels of intelligence. The future development trend may be to seek a balance between using local models to ensure data security and rapid response, while relying on cloud models to solve complex problems.