Alibaba Advances in Generative AI with New Region-Specific Models

Chinese e-commerce giant Alibaba has recently made significant advancements in the field of generative artificial intelligence (AI), with a particular focus on developing new models tailored to the Southeast Asian market and new models aimed to optimize online shopping experience at its Taobao Tmall Commerce Group.

This week, Alibaba Group’s research institute, Damo Academy, launched the first artificial intelligence large model version trained on Southeast Asian languages named SeaLLM, as well as a chatbot named SeaLLM-chat, reflecting Alibaba’s emphasis on the Southeast Asian market.

According to information released on Alibaba’s website Alizila, the AI large language models released this time include two different versions with 13 billion and 7 billion parameters, capable of processing languages including Vietnamese, Indonesian, Thai, Malay, Khmer, Lao, Tagalog, and Burmese.

At the same time, the Taobao Tmall Commerce Group has just restructured its AI business, consolidating it from about twenty teams to four, and has internally released its own large model product, ‘Turing’. The AI team of the International Digital Commerce Group now has over a hundred members.

According to insiders, this large model product will not be released to the public. Internally, it is mainly used for two purposes: one is for search, advertising, and recommendations, and the other is for content creation in the browsing experience within the Taobao Tmall Commerce Group.

Earlier in November, Alibaba’s Intelligent Information Business Group released the Quark large model, which is a full-stack self-developed model with hundreds of billions of parameters.

According to the introduction, the Quark large model is an application-oriented large model designed for search, productivity tools, and asset management assistance.

In search applications, it aims to broaden application scenarios and enhance user experience through multi-modal understanding of images and text, generation of professional knowledge, and innovation in interaction methods.

This week, the self-developed Quark large model was officially registered with Chinese regulators.

Alibaba stated that the large model based on Southeast Asian languages was launched against the backdrop of growing demand for large models in Southeast Asian countries. The aim is to create AI large models that are more inclusive and regionally relevant, thereby reflecting the subtle differences in Southeast Asian cultures.

Currently, most of the AI large models in the Southeast Asian region come from Western countries, and the training of these models primarily uses English or other Latin-based languages.

The SeaLLM large model is trained through a variety of Southeast Asian languages, has undergone special vocabulary processing for non-Latin languages. After tokenization of sentences of the same length, the sequence length obtained is one-ninth of that of ChatGPT, and it possesses more complex task execution capabilities.

Alibaba has long-term cooperation with Nanyang Technological University in the field of multilingual artificial intelligence research.

Luu Anh Tuan, Assistant Professor at the School of Computer Science and Engineering (SCSE) at Nanyang Technological University, stated that the launch of the SeaLLM large model could bring new opportunities for millions of people who speak languages other than English and Chinese.

In April, Alibaba released the Qwen-72B model, which was one of the largest models open-sourced in China at that time.


Related News: