Written by｜Zhao Guangli
Have you heard of crowdfunding for a “large model”? Just like crowdfunding for cars or houses?
This actually happened not long ago.
A “Crowdfunding” Story
In May 2021, nearly a thousand scientists and volunteers from different nationalities and professional fields initiated a crowdfunding project for a large model. Why such a large mobilization? It starts with the application of large models.
The full name of the large model is “Artificial Intelligence Pre-trained Large Model”. It is called a “large” model because, compared to ordinary AI models, it has massive training data and an extremely large number of parameters, allowing it to handle tasks in various scenarios.
If developing an AI model is like cooking a dish, an AI large model is like providing “pre-made meals” that can be simply heated and eaten. This greatly saves time and reduces the repetitive labor in the development process of AI models from 0 to 1.
However, due to the high cost of training large models, most AI large models are controlled by large tech companies or specialized institutions, limiting access for ordinary scholars and developers. Thus, these scientists thought of crowdfunding to collaboratively create an open-source AI large model for easier use and scientific research.
This endeavor was indeed successful. In just over a year, the project received approximately $7 million in public funding, creating a multilingual model with 176 billion parameters, comparable to the well-known GPT-3. This large model is called “BLOOM”. It is reported that from code to dataset, BLOOM is fully open to the public, and everyone can download and use it.
However, it is understood that downloading and using BLOOM requires certain local hardware capabilities, so currently, BLOOM is only available for some large research teams. Moreover, from the perspective of the BLOOM dataset, although the Chinese dataset accounts for a significant proportion, it only constitutes about 16.25% (including traditional Chinese), which is not very convenient for Chinese scientists and developers to use.
Chinese AI researchers have the same desire for large models, which further drives the research and application of AI in Chinese. Especially as large models have become more numerous and their parameter scales have grown larger over the past two years, they seem increasingly esoteric and less accessible from the outside. Can these large models be made open-source?
No need for crowdfunding or waiting, with the debut of “Magic Model” at the 2022 Yunqi Conference, this day has arrived.
Alibaba DAMO Academy Shares Generously
On November 3, 2022, good news came from the Yunqi Conference in Hangzhou: Alibaba DAMO Academy, in collaboration with the CCF Open Source Development Committee, launched the AI model community “Magic Model” (ModelScope), aiming to lower the barrier to AI applications. As the initiator, DAMO Academy first contributed over 300 verified high-quality AI models to the Magic Model community, with more than one-third being Chinese models, all fully open-source and transformed into directly usable services.
For many AI researchers, developers, and enthusiasts, this is like receiving a pillow just as they were about to doze off.
To build a strong Magic Model community, DAMO Academy’s initial contribution of over 300 models includes over 150 industry-leading models across various intelligent fields such as natural language processing, vision, speech, and multimodal, many of which are pre-trained multimodal large models, including the previously announced Tongyi large model series by DAMO Academy.
Alibaba DAMO Academy’s actions are indeed sincere. So far, globally, no other institution or tech company has offered such a large-scale large model for free open use. As Alibaba Group’s Senior Vice President and DAMO Academy Vice President Zhou Jingren said, “DAMO Academy has shared everything this time.”
“If we hold back today, this project cannot be completed,” Zhou Jingren stated in an interview. “We hope to set a good precedent by making our best models available.”
DAMO Academy is not working alone to build the Magic Model community. Among the first batch of cooperative institutions in the community are DeepMind Technology, LanZhou Technology, Zhiyuan AI, and the University of Science and Technology of China.
Correspondingly, DeepMind Technology’s protein sequence prediction model Uni Fold Monomer, LanZhou Technology’s Mencius series language large model, and Zhiyuan AI’s multilingual pre-trained model have also “joined” the Magic Model community.
In addition, the University of Science and Technology of China and Zhejiang University are also exploring a series of collaborations with the Magic Model community in AI talent cultivation. Zhou Jingren expressed a strong desire to collaborate with domestic scholars and developers to co-build the community.
It is reported that the models open to the Magic Model community are compatible with various mainstream AI frameworks, supporting multiple training and service deployment methods, allowing users to choose according to their preferences. Furthermore, the community is open to all developers and will implement council management, aiming to promote large-scale AI applications without profit as a goal.
This kind of open-source community for sharing and advancing AI models means, for developers and enthusiasts, “the flowers should be picked when they bloom,” and the feedback received during the evaluation phase indeed reflects this.
“There are simply too many models in the community, it’s like a blessing for student developers!” One student developer who participated in the evaluation phase of the Magic Model community felt that he was born at the right time: “Some of the content we learn in class often leaves us at a loss when practicing outside of class, and the teacher told us to try this new community.”
In the Magic Model community, he first searched for and used the vocoder model HIFI-GAN, which is a training model with a wide range of applications suitable for Chinese speech synthesis scenarios such as dubbing, virtual hosts, and digital humans. During the process, he also experienced the community’s strength: “It can directly point out the mistakes I made in my operations, saving me a lot of time searching for errors and bugs.”
Community Under the MaaS Concept
Zhou Jingren proposed that in the development and application of AI technology, models serve as a carrier. “High usage barriers limit the potential of AI.” To accelerate breakthroughs in AI application challenges, Alibaba DAMO Academy believes that a corresponding service system should be built around models. Based on open-source large models, they proposed the concept of “Model as a Service” (MaaS).
The core idea of MaaS is to provide various services around the model, from providing models to offering a wide range of services. Zhou Jingren stated that the biggest challenge in building an open-source community is to get more people involved in community construction, enabling more developers to solve practical problems through the community and actively use and provide feedback. Therefore, it is essential to focus on “community friendliness” based on the MaaS concept.
Thus, they are taking action. First of all, the Magic Model community has an abundant supply of Chinese AI models, with over 100 Chinese models currently available to better meet local demands; secondly, the Magic Model community provides an easy-to-use model usage platform, making it no longer difficult to run AI models—what used to take days for code downloads, secondary development, installation, and verification now only takes a few hours or even minutes.
Additionally, through newly developed calling interfaces and unified configuration files, the platform provides one-stop services for model exploration, environment installation, inference verification, and training optimization. Users can experience model effects online with zero code, achieve model inference with one line of code, and model tuning and customization with ten lines of code. Meanwhile, the platform also offers online development capabilities and computing power support, requiring no installation or deployment—just open a webpage to develop AI models.
“Magic Model (ModelScope) is a community built under the MaaS concept,” Zhou Jingren believes that the MaaS concept establishes a model-centered, full lifecycle management mechanism, which means providing developers with a series of support. Only in this way can models be quickly transformed from a development environment into a production environment and quickly link models with business scenarios interactively.
“MaaS is an important direction for the future development of artificial intelligence, and if practiced well, it will promote significant growth in the AI industry,” Zhou Jingren said.
Not a Milestone, Just a Starting Point
At the 2022 Yunqi Conference, Chinese Academy of Sciences Academician and Director of the CCF Open Source Development Committee Wang Huaimin expressed hope that the open-source of AI models could build a “national library” through the joint efforts of the market, society, and government.
He stated that open-source is an important driving force for AI development, and the Magic Model community, as a new type of AI open-source community, will not only strongly promote the widespread application of AI but also help China grow from a participant in the open-source world to a leader.
Chinese Academy of Sciences Academician and Peking University Professor E Wei Nan believes that the Magic Model community is an important attempt to provide general research tools for the new research paradigm of AI for Science, facilitating the transition of scientific research from a “small workshop” model to an “Android” model, avoiding repetitive work and forcing original innovative research.
Compared to the crowdfunding behavior initiated by foreign scholars at the beginning of the article, the emergence of the Magic Model community undoubtedly marks a significant event in the field of Chinese AI industry. However, even with this excitement, Zhou Jingren emphasized that the community’s launch is not a “milestone” but rather “a starting point.”
“Because today we need to jointly build a model ecosystem based on MaaS, continuously enriching and improving model services.” Zhou Jingren revealed that based on the current trend, he expects new models to be launched in the Magic Model community every month, and the number of models will soon double, covering all aspects of various application fields.
“We are just getting started and will continue to release new models; we hope more developers will contribute to make the community vibrant and the models ‘play’ effectively, thereby unleashing the infinite potential of large AI models,” Zhou Jingren said.
Magic Model community address: modelscope.cn（Original text can be accessed directly）
Editor | Fang Yuan
Layout | Hua Yuan
Cooperation Inquiries:[email protected]
Submission Inquiries:[email protected]