ANT GROUP CO., LTD. uses domestically produced AI Chip to train large models, which can further reduce costs.
Recently, the ANT GROUP CO., LTD. Ling team published a technical achievement paper. The paper shows that ANT GROUP CO., LTD. has launched two different scales of MoE large language models - Ling-Lite and Ling-Plus. The former has a parameter scale of 16.8 billion (with 2.75 billion active parameters), while the Plus base model has a parameter scale of up to 290 billion (with 28.8 billion active parameters), and both achieve industry-leading performance.
In addition to the self-developed large model with leading performance, the biggest breakthrough of this technical paper is the proposal of a series of innovative methods to enhance the efficiency and accessibility of AI development in resource-constrained environments. Experiments show that its 300 billion parameter MoE (Mixture of Experts) large model can achieve efficient training on low-performance devices using domestically produced GPUs, with performance comparable to fully using.$NVIDIA (NVDA.US)$Chip, the same scale dense model, and MoE model.
Self-developed large model efficiently trained on low-performance hardware.
Currently, the technical achievement paper by ANT GROUP CO., LTD. Ling team "Every FLOP Matters: Scaling 300 billion parameter Mixture of Experts LING Large Model Without Advanced GPUs" has been published on the preprint platform Arxiv.
According to the technical achievement paper, although a series of MoE large models like DeepSeek, Alibaba Tongyi Qianwen, and MiniMax have demonstrated outstanding performance in specific tasks, the training of MoE models usually relies on high-performance computing resources (such as NVIDIA H100/H800 and other advanced GPUs), and the high costs limit their widespread application in resource-constrained environments. Meanwhile, in recent years, the shortage of NVIDIA high-performance chips continues, while low-performance accelerators are more abundant and have lower single-machine costs. This disparity highlights the necessity of constructing a seamless switching technology framework for heterogeneous computing units and distributed clusters.
Therefore, the goal set by the Ling team is to "expand the model without using advanced GPUs" and aims to break through resource and budget constraints to achieve efficient large language model training through innovative training strategies, thereby promoting the democratization of AI technology.
Specifically, the innovative strategies proposed by the team include: 1) Architecture and training strategy innovation: dynamic parameter allocation and mixed precision scheduling technology; 2) Upgrade of training abnormal handling mechanism: adaptive fault tolerance recovery system shortens interruption response time; 3) Optimization of model evaluation process: automated evaluation framework compresses validation cycle by over 50%; 4) Breakthrough in tool invocation capability: instruction fine-tuning based on knowledge graph improves the execution accuracy of complex tasks.
According to the technical paper, the Ling team pre-trained Ling-Plus on five different hardware configurations with 9 trillion tokens, where the pre-training cost for 1 trillion tokens using high-performance hardware configuration was about 6.35 million yuan, but after applying ANT GROUP CO., LTD.'s optimization method, the training cost using low-spec hardware would drop to around 5.08 million yuan, saving nearly 20%, achieving performance comparable to Alibaba Tongyi Qwen2.5-72B-Instruct and DeepSeek-V2.5-1210-Chat.
Previously, DeepSeek developed V3 and R1 models with performance comparable to top models through a series of algorithm innovations and engineering optimizations using the lower-performance NVIDIA H800, opening up new avenues for training large models and allowing more enterprises and research Institutions to see the possibility of reducing costs and improving efficiency. If the technological achievements of ANT GROUP CO., LTD. are validated and promoted, it means that domestically produced large models can seek lower-cost and more efficient domestic chips or other alternatives to further reduce dependence on NVIDIA chips.
ANT GROUP CO., LTD. continues to increase investment in AI applications and humanoid robots.
According to reporters, the BaiLing large model, as ANT GROUP CO., LTD.'s self-developed large model, focuses on applications in daily services, financial services, and medical health scenarios. In May last year, ANT showcased multiple AI innovative application products during an open day, and for the first time announced the AI application matrix. ANT GROUP CO., LTD.'s CTO He Zhengyu revealed that three applications based on ANT BaiLing large model are the key breaking directions for ANT at present: lifestyle manager, medical assistant, and financial assistant.
On March 21, ANT announced the latest progress in the AI medical field: the release of an AI product system upgrade aimed at medical institutions, doctors, and users, among which, for medical institutions, we collaborated with the Huawei medical and health team, Alibaba Cloud, Apple, etc. to launch the "ANT Medical Large Model Integrated Machine" full-stack solution; for doctors, we released a series of AI doctor assistant tools; at the same time, the health application "AI Health Manager" for users also launched more than ten new features such as intelligent thinking and health self-testing.
In addition to AI, ANT GROUP CO., LTD. has also been very active in the humanoid robot field recently. In February of this year, a recruitment platform showed that ANT GROUP CO., LTD. opened recruiting for positions related to embodied intelligent humanoid robot systems and applications, with an annual salary reaching over one million. As early as December last year, ANT GROUP CO., LTD. established Shanghai Ant Lingbo Technology Co., Ltd., focusing on research and development of embodied intelligence technology and products.
According to news from Pudong, on March 11, Shanghai ANT GROUP CO., LTD. held an unveiling ceremony in Pudong, Shanghai. It was introduced that ANT GROUP CO., LTD. is the main carrier for expanding the Asia Vets and Siasun Robot&Automation business, dedicated to creating industry-leading robot products in the fields of home, Retirement, and Medical health. This landing will work hand in hand with Pudong to promote mutual development on the future new track of Industries, helping Pudong accelerate technology leadership, industry aggregation, and industrial upgrading, creating an innovative highland led by humanoid robots and an innovation industry ecosystem with industry influence.
As an emerging field, Asia Vets humanoid robots possess broad market prospects and immense commercial potential. ANT GROUP CO., LTD. may aim to explore new growth points by entering this field and drive new experiences in human-computer interaction. Furthermore, Asia Vets humanoid robot technology can generate synergy with ANT GROUP CO., LTD.'s existing CNI Xiangmi Lake Fintech Index business, jointly promoting technological innovation and business upgrades.
Industry insiders analyze that Asia Vets humanoid robots are an emerging field with broad market prospects. ANT GROUP CO., LTD.'s layout in this area helps explore new growth points and, based on its own technological advantages and existing business layouts in AI, Big Data, and Cloud Computing, accelerate the research and development of humanoid robots and their applications in related business scenarios.
Is investing always stepping on a landmine?Futubull AIGoing live! Accurate answers, comprehensive insights, seizing key opportunities!
Editor/rice