Alibaba Cloud debuts Model Studio

Alibaba Cloud has launched an artificial intelligence (AI) model studio that provides access to foundation models and model training tools to speed up deployment of large language models (LLMs).

The models available through the Alibaba Cloud Model Studio includes the company’s Qwen models, including the Qwen-72B and Qwen-1.8B, the 72-billion-parameter and 1.8-billion-parameter versions of its proprietary foundation model, Tongyi Qianwen, as well as third-party and industry-specific models.

Speaking to Computer Weekly on the sidelines of a recent Alibaba Cloud event in Singapore, Guo Dongliang, vice-president for product and solution at Alibaba Cloud International, said Model Studio marks an effort by Alibaba Cloud to help customers avoid the challenges that come with building and deploying LLMs.

“We’ve gone through the tough journey of using and building models for ourselves and have overcome lots of technical obstacles,” he said. “Our customers may have to go through the same journey and to help them to avoid that, we’ve embedded all our experience into our products.”

Guo said Model Studio is aimed at large enterprises that have strict security and privacy requirements, and would like to run and manage LLMs in a controlled environment. With Model Studio, they can monitor and identify risky content, and filter or block undesirable information based on responsible AI principles.

Organisations can also train foundation models in Model Studio by creating, labelling and managing training datasets, customise model training with adjustable parameters, as well as evaluate and deploy foundation models easily.

To assuage the security concerns of customers, Guo said Alibaba Cloud has built guardrails into all its products. For example, he said, outputs from LLMs are subject to security policies and aligned with the company’s ethics guidelines.

“We also make sure that customer data is within the control of our customers and this not just for GenAI but also for Alibaba Cloud as a whole. We will continue to do so for Model Studio and other future products as well,” he added.

As the first global hyperscaler to establish local datacentres in key emerging Southeast Asian markets such as Malaysia and Indonesia, Alibaba Cloud has also built a family of LLMs catered to local languages such as Bahasa Indonesia.

Dubbed SeaLLMs, the models, which were built on Meta’s Llama 2, were designed to cater to the linguistic diversity of Southeast Asia, enabling businesses to leverage chatbots that not only comprehend, but also reflect the social norms, customs, stylistic preferences and legal considerations in the region.

However, Guo revealed that Alibaba Cloud’s Qwen model, which was trained from scratch using data that adheres to laws and compliance requirements, will be a more powerful foundation model for Southeast Asia for local languages, noting that a Japanese partner had trained it on the Japanese language to deliver better performance than Llama 2.

Selina Yuan, president of Alibaba Cloud’s international business, expects the surge in demand for AI capabilities to fuel the company’s growth in 2024. “With our tremendous capacity in cloud computing and LLMs, we expect to serve more customers in local markets,” she said.