According to the Star Daily on March 24 (Reporter Huang Xinyi), following Alibaba CEO Wu Yongming's announcement to fully "AI-enable," related actions by ANT GROUP CO., LTD. in AI have also been ongoing recently.

Today, in response to reports on the training costs of the Ant Bai Ling large model, ANT GROUP CO., LTD. promptly responded to the Star Daily, stating: Ant continues to optimize for different chips to reduce the cost of AI applications, and has made certain progress, which will gradually be shared through open-source.

The latest research paper released by ANT GROUP CO., LTD. this month shows that it has launched two large MoE language models of different scales—Bai Ling Lite and Bai Ling Plus. The former has a parameter scale of 16.8 billion (with 2.75 billion active parameters), while the Plus base model's scale reaches an impressive 290 billion (with 28.8 billion active parameters). Experiments indicate that its 300 billion parameter MoE large model can achieve efficient training on low-performance devices using domestically produced GPUs, performing comparably to the fully NVIDIA chip-based dense models and MoE models of the same scale.

According to the paper, while series of MoE large models like DeepSeek, Alibaba Tongyi Qianwen, and MiniMax exhibit exceptional performance in specific tasks, the training of MoE models typically relies on high-performance computing resources (such as advanced GPUs like NVIDIA H100/H800), which significantly limits their widespread application in resource-constrained environments due to high costs. Additionally, the continuous shortage of high-performance NVIDIA chips in recent years has made low-performance accelerators more readily available and lower in cost per unit, highlighting the necessity of constructing a seamless switching technology framework for cross-heterogeneous computing units and distributed clusters.

Therefore, the goal set by the Ant team is to "expand the model without using advanced GPUs," aiming to break through resource and budget constraints to achieve efficient training of large language models through optimizations and implementations in model training environment, optimization strategy, infrastructure, training process, evaluation results, and inference.

The Ant Ling team has pre-trained Ling-Plus on 9 trillion tokens across five different hardware configurations. Among them, the pre-training cost for training 1 trillion tokens with a high-performance hardware configuration was approximately 6.35 million yuan, but Ant's optimization methods will reduce the training cost when using low-spec hardware to around 5.08 million yuan, saving nearly 20% in costs, ultimately achieving performance comparable to that of Alibaba Tongyi Qwen 2.5-72B-Instruct and DeepSeek-V2.5-1210-Chat.

As a large model developed by ANT GROUP CO., LTD., the Bailing large model focuses on applications in life services, financial services, medical health, and other scenarios. In the future, ANT GROUP CO., LTD. plans to open source the ANT Bailing large model Ling-Plus and Ling-Lite.

Recently, ANT GROUP CO., LTD. has been frequently increasing its investment in the field of AI, with medical being a major focus area. On March 21, ANT GROUP CO., LTD. announced the latest AI product system upgrades for medical institutions, doctors, and users. For medical institutions, it launched the "ANT Medical Large Model Integrated Machine" in cooperation with Huawei Medical Health Corps and Alibaba Cloud; for the 0.29 million registered doctors on Haodaifu Online, it released a series of AI doctor assistant tools; at the same time, the health application "AI Health Butler" introduced more than ten new features such as intelligent thinking and health self-assessment for user service.

In addition, ANT GROUP CO., LTD. is also expanding directions such as embodied intelligence and AI glasses.

ANT GROUP CO., LTD. has registered the Shanghai Ant Lingbo Technology Co., Ltd. As the main carrier for ANT GROUP CO., LTD. to expand embodied intelligence and robotics business, Ant Lingbo Technology will focus on family, retirement, medical health, and other fields, assisting Shanghai Pudong in accelerating technology leadership, industry aggregation, and industrial upgrading, creating an innovative highland for embodied intelligence led by humanoid robots and an innovative industrial ecosystem with industry influence.

Job postings show that ANT GROUP CO., LTD. is currently recruiting product experts for AI smart glasses, requiring experience with 2C products. "Star Daily" learned from sources close to ANT GROUP CO., LTD. that ANT is indeed expanding its preparations for smart glasses-related business.

Editor/Jeffy