MiniMax Vice President Liu Hua: Standardized large-scale models are the future, and technological evolution is superior to customization | Frontline

Text | Tian Zhe

Editor | Su Jianxun

As artificial intelligence enters the era of large-scale models, the global technological competition is gradually focusing on China and the United States. The high investment in computing power and the massive demand for data determine that the research and development threshold of large-scale models is extremely high, and the large-scale model manufacturers in a few leading countries are competing for cutting-edge advantages in this technological race.

MiniMax is one of the early domestic large-scale model manufacturers to go global. It is reported that MiniMax's overseas revenue this year will exceed 70 million US dollars.

Currently, MiniMax has launched two C-end products, HailuoAI and Talkie, in the overseas market, focusing on AI video/audio generation and AI content community respectively. 36Kr has learned that in addition to the relatively large proportion of the overseas C-end AI application Talkie, the Hailuo AI subscription service and the B-end API service of MiniMax have also improved.

Recently, Tian Feng, the General Manager of Cloud Native in the Northern Region of Tencent Cloud, and Liu Hua, the Vice President of MiniMax, had a conversation with 36Kr and other media. From the global competition pattern, technical path to the commercialization direction, they jointly discussed the development trend of large-scale model technology and the collaborative value between them.

Liu Hua introduced that in addition to the existing paid subscription system, the advertising business is also one of the commercial revenue sources of Talkie. In addition, MiniMax will not develop customized model projects in the overseas market for the time being, and currently still serves B-end customers with standardized API interface services.

The reason is that start-up companies should invest their main resources in the research and development and iteration of core technologies, rather than consuming a lot of energy in customized projects that meet individual needs. "If a large-scale model requires a lot of customized development to meet customer needs, it indicates that this generation of models is not yet mature. Instead of investing time and cost in customization, it is better to accelerate the model iteration," Liu Hua said.

He revealed that MiniMax has determined the future research and development route, mainly focusing on reducing the model error rate, achieving infinitely long input and output, and developing in the multi-modal route.

First, in terms of the model error rate, he believes that the error rate of the previous generation of GPT series models is about 30%. The reason why it has not attracted much attention from users is that it is mostly used in the cultural and creative field where there is no standard answer. If large-scale models are used in scenarios such as production, research and development, and scientific research design, it is best to reduce the error rate to a single digit. And the ideal error rate of MiniMax is 2%-3%.

Secondly, as the tasks of large-scale models gradually expand from text to voice and video, the amount of Tokens required is also increasing rapidly. Therefore, the key to the landing of new technologies lies in improving the ability of large-scale models to handle large-scale input and output.

Currently, the latest Abab 7 series model developed by MiniMax is based on a new architecture of MoE and Linear Attention mechanism, which can significantly reduce the computational complexity of long texts.

According to the observation of 36Kr, the large-scale model industry in the United States still maintains a certain degree of leadership. Industry-leading manufacturers such as OpenAI, Anthropic, and XAI occupy strong resource and technical advantages. For example, the company size of OpenAI is close to 2,000 people, and the annual computing power cost may reach tens of billions of US dollars.

However, Chinese large-scale model manufacturers are catching up very quickly, especially in the fields of voice and video. China's voice API service has approached the level of GPT-4o. If well-known domestic technology companies also enter this track, they will inevitably occupy an important position with their resource advantages.

Liu Hua believes that Chinese AI start-up companies still have a broad space for development. Just as MiniMax has received support from cloud manufacturers such as Tencent Cloud, and at the same time has achieved a capital cycle through successful commercialization.

In addition, Liu Hua believes that if it is necessary to better serve users at home and abroad, it is necessary to insist on independent research and development to create a truly native solution, so as to have a large-scale model with independent competitiveness.

When it comes to the cooperation with MiniMax, Tian Feng, the General Manager of Cloud Native in the Northern Region of Tencent Cloud, introduced that Tencent Cloud provides MiniMax with a series of high-performance intelligent products integrating computing, storage, and network, allowing MiniMax to focus more on the training and engineering of the model itself. Among them, the object storage product provides a metadata acceleration solution to ensure performance, as well as a variety of refined management measures for data governance to reduce costs and increase efficiency. The data lake product is specially optimized for the preprocessing of corpus data to improve task processing performance, helping MiniMax save more than 30% of computing power and improve performance by more than 35.5%.

It is reported that Tencent Cloud's high-performance computing cluster can detect problems within one minute, locate problems within three minutes, and restore the system within the fastest five minutes through a systematic operation and maintenance mechanism. The daily failure number of its thousand-card cluster has been refreshed to 0.16, which is one-third of the industry average level.

At the same time, Tencent Cloud's Xingmai Network can support the training of large-scale models, achieving that the network communication accounts for as low as 6% (communication time accounts for the proportion of the total time) in the training process of large-scale models, and the overall training efficiency is improved by more than 20%.

Just as American large-scale model manufacturers have formed a head effect, this phenomenon will also appear in the Chinese market. Liu Hua believes that in the future, only a few enterprises will develop basic large-scale models, and most enterprises will gradually shift to the research and development of the AI application layer.

As for MiniMax, it will still insist on investing in the research and development of core technologies, using good technologies to drive good products, and then letting good products bring good service experiences and word-of-mouth to feed back to the technology.