Ali's large model "goes to sea" to compete for the Southeast Asian market

The DAMO Academy, a research arm of Alibaba Group, this week unveiled SeaLLM, the first version of a large AI model based on Southeast Asian language training, and a chatbot called SeaLLM-chat, reflecting Alibaba's focus on the Southeast Asian market.

According to information released by Alibaba.com Alizila, the AI large language model released this time includes two different versions of 13 billion parameters and 7 billion parameters, and the languages that can handle include Vietnamese, Indonesian, Thai, Malay, Khmer, Lao, Tagalog, and Burmese.

Large model “ Sea ” Go with the flow

Alibaba said that the large model based on the Southeast Asian language version was launched against the backdrop of growing demand for large models in Southeast Asian countries, and aims to reflect the nuances of Southeast Asian culture by creating more inclusive and regionally relevant AI large models.

At present, most of the large AI models in Southeast Asia come from Western countries, and the training of large models is mainly in English or other Latin languages. "This innovation will accelerate the adoption of AI and support historically underrepresented communities in the digital space," said Lidong Bing, Director of the Language Technology Lab at Alibaba DAMO Academy. ”

According to reports, the SeaLLM large model has been trained in a variety of Southeast Asian languages, and its text interpretation and processing time for non-Latin languages is nine times longer than that of models such as ChatGPT, and it has more complex task execution capabilities. It outperforms most open-source large models in understanding a variety of disciplines from science, chemistry, physics, to economics using Southeast Asian languages.

Alibaba has a long-standing collaboration with Nanyang Technological University on multilingual AI research. Luu Anh Tuan, an assistant professor at Nanyang Technological University's School of Computer Science and Engineering (SCSE), said the launch of the SeaLLM model could bring new opportunities to millions of speakers of languages other than English and Chinese, and could also help companies get more involved in the Southeast Asian market and build their own chatbot assistants.

Judging from the layout of Alibaba's large model, the first financial reporter noticed that different business units under the group are developing large models. Recently, it was reported that Taobao Tmall Business Group is also preparing to build a large model team and has begun recruitment. The first financial reporter asked the relevant person in charge of Taobao Tmall Business Group to confirm the news, but the other party did not comment.

At present, Alibaba's most extensive large model is Tongyi Qianwen, a subsidiary of Alibaba Cloud. Just earlier this month, Alibaba Cloud open-sourced the 72 billion parameter and 1.8 billion parameter versions of the model. Since August this year, the model has been downloaded more than 1.5 million times on the Model Scope and Hugging Face.

The "going overseas" of Alibaba's large model this time is also an inevitability to follow the business development of Alibaba Cloud. Alibaba Cloud was the first to deploy in Southeast Asia. According to Gartner, Alibaba Cloud's market share in Malaysia and Indonesia will be close to 30% in 2021, becoming one of the important engines supporting growth.

Currently, Alibaba Cloud has opened data centers in Singapore, Indonesia, the Philippines, Malaysia, and Thailand, with a total of 10 availability zones in the Southeast Asia region. In September last year, Alibaba Cloud announced that it would invest 7 billion yuan in the next three years to build an international localization ecosystem. Alibaba Cloud's customers include Lazada, a Southeast Asian e-commerce company controlled by Alibaba Holdings, REDMART, a Singaporean retail company, and Yippi, a Malaysian social platform.

Three forces compete for the large model market in Southeast Asia

Nowadays, China's large-scale model "going overseas" is also aimed at the needs of Chinese e-commerce and logistics enterprises to deepen their layout in Southeast Asia.

Southeast Asia is becoming a new hot spot for Chinese e-commerce competition. For example, Alibaba's Southeast Asian e-commerce platform, Lazada, aims to achieve a turnover of $100 billion by 2030, and the platform has 300 million consumers in the region.

On December 11, TikTok announced that it had reached an agreement with Indonesian technology giant GoTo, and the Indonesian business of TikTok e-commerce platform TikTok Shop and GoTo's e-commerce platform Tokopedia merged to form PT Tokopedia, which TikTok has a controlling stake in.

Tokopedia is also known as the "Indonesian Taobao", which marks the return of TikTok e-commerce to Indonesia. The platform will also be Lazada's biggest competitor in Southeast Asia. According to data from Singapore-based consulting firm Moten Ventures, in 2022, Indonesia's e-commerce transaction volume was about US$52 billion, of which about US$2.5 billion came from TikTok, accounting for 57% of TikTok's GMV (gross merchandise value) in Southeast Asia.

Another big demand is live streaming. With TikTok and other cross-border e-commerce businesses in Southeast Asia, the recruitment and management of foreign language anchors are facing challenges, and cross-border e-commerce hopes to use the power of generative AI and other technologies to quickly and cost-effectively deploy live broadcast e-commerce and bring business increments through AI digital human live broadcasts and other methods.

"In the retail field, the main application scenarios of generative AI include inventory management forecasting, personalized recommendations, customer service automation, market trend forecasting, virtual fitting and product display," Lu Yanxia, research director of IDC China, told CBN. ”

Alibaba's bid to compete in Southeast Asia's large-scale model market also reflects the fact that Chinese companies are taking advantage of language to seek differentiated competition in the global and domestic large-scale model boom. According to incomplete statistics, as of July this year, Chinese companies and research institutions have released more than 130 large language models, setting off a "100-model war" in China.

Despite the numerous models, analysts say Chinese AI companies still face multiple challenges in a rapidly changing market, such as how to attract more users and lack of killer apps.

In addition to Alibaba, the National Research Fund for Singapore's National Research Fund (NLP) has launched the AI Multimodal Large Model Development Program (NMLP), for which the Singapore government will invest a total of US$52 million.

The first financial reporter also paid attention to the fact that in June this year, iFLYTEK also announced its strategy for large models in Southeast Asia. iFLYTEK said that Southeast Asia has become the first stop and strategic center of iFLYTEK's overseas business.

Zhang Junyi, managing partner of Aowei Consulting, told the first financial reporter: "The Southeast Asian market is mainly composed of three forces, one is the local force in Southeast Asia, such as the large model supported by the Singapore state, the second is the Chinese force, represented by Alibaba, and the third force is the large model from European and American companies such as OpenAI, which is also the most used at present, and the 'Three Kingdoms Competition' will continue in the future." ”