Alibaba Cloud announced the third round of price reductions for large models this year, with the Tongyi Qianwen visual understanding model seeing price reductions of over 80% across the board.

Among them, the Qwen-VL-Plus saw a direct reduction of 81%, with an input price of only 0.0015 yuan per 1,000 tokens, setting the lowest price on the internet; the higher-performing Qwen-VL-Max dropped to 0.003 yuan per 1,000 tokens, a decrease of up to 85%. According to the latest prices, 1 yuan can process approximately 600 images at 720P or 1,700 images at 480P.

This undoubtedly counters the launch of ByteDance's Doubao visual understanding model. At the Volcano Engine Force conference held on December 18, the input price for the Doubao visual understanding model released by ByteDance was 0.03 yuan per 1,000 tokens, enabling the processing of 284 images at 720P for 1 yuan. At that time, ByteDance claimed that this price was 85% cheaper than the industry average.

Since the rise of large models, related companies have been burning money with significant investments. With the multiple rounds of price cuts for large models, startup companies with weaker financial strength face greater impacts. ByteDance's late breakthrough in the large model field also poses considerable pressure on internet companies like Baidu and Alibaba.

Price cuts and traffic investment, the large model cash-burning war continues.

As early as May 2024, companies like Zhizhu AI, ByteDance, and Alibaba Cloud had already started a wave of price cuts for large models, reducing the price of large model inference computing power by 90%. This led to large model pricing entering the "cent era."

In December 2024, the dark side of the moon Kimi, ByteDance's Doubao, and Alibaba Cloud Tongyi Qianwen will successively launch new visual understanding large models and start a new round of price cuts. People involved in large models told the Star Daily reporter that the price war initiated by major industry players first impacts startup companies. "Especially if these companies want to develop B-end business, it is very difficult to compete with the large companies."

A responsible person from a large model startup stated to the Star reporter that competition isn't necessarily a bad thing, but if it simply turns into a price war, then that is certainly not good. "If everyone is competing on technological research and development capabilities, providing customers with better cost-effectiveness and better service, that is a good thing, and it is precisely this kind of competition that allows Chinese companies to have an advantage in overseas markets."

In addition to capturing the market through price reductions, large model companies are also trying to attract new users with huge advertising investments.

According to statistics from the Penghua CSI Mobile Internet Index Fund(LOF)-A marketing platform AppGrowing, Kimi's advertising spending in the third quarter was approximately 0.15 billion yuan, Doubao around 0.2 billion yuan, and Yuanbao about 80 million yuan. Doubao was 0.2 billion yuan. However, an individual related to AI large models informed reporters that this data is not accurate. Nonetheless, it is an undisputed fact that both Kimi and Doubao are significantly increasing their large model advertising investments. Ads for AI products like Kimi, Doubao, and Zhibu can frequently be seen on platforms such as Douyin, Kuaishou, Bilibili, Xiaohongshu, Douban, as well as offline in subways and airports.

A large amount of advertising investment has allowed Doubao, Kimi, and others to reach more users. In the November global monthly active ranking of the "AI Product List," Doubao App's monthly active users reached 59.98 million, second only to OpenAI's ChatGPT, ranking second globally. However, in February 2024, Doubao's monthly active user count was only 1.73 million.

Some industry analysts believe that the cash-burning war of large model investment is very similar to the early days of the mobile Internet, where various applications burned money on promotions and even offered direct subsidies to compete for users. However, the hundreds of millions in advertising costs may not necessarily translate into long-term consumer willingness, and retention remains an unknown factor for large model manufacturers.

Yang Zhiling, founder of the Dark Side of the Moon, stated in an interview with the Star Daily and others that the most critical task is to improve user retention, which is currently Kimi's core mission, with no upper limit on restrictions.

As a new generation of technology, large models should focus more on technological research and development, said a person from the AI large model team to the Star Daily reporter. "Using better technology to develop products for everyone's use. If everyone focuses on traffic generation, it may lead to homogenized technology, with not much difference in technology. The technology for large models is still rapidly developing, so the best way is to develop better and higher-level large model technologies, and then based on that, develop products with more performance, or more innovative products. I think that is a better approach."

Alibaba and Baidu need to break through at the consumer end.

In order to seek a way out amidst competition from large companies, many AI startups have adjusted their positioning.

For example, Baichuan Intelligence is betting on the AI medical health field. On December 31, 2024, Baichuan Intelligence announced a strategic cooperation with the Haidian District Health Commission in Peking to jointly create a grassroots medical service model of "AI doctor - real doctor dual medical collaboration."

Going overseas is also an important path to seek commercial closure, and MiniMax is a standout in this regard. The Star Daily reporter learned from informed sources that the annual revenue of applications under MiniMax is expected to exceed 70 million dollars in 2024. The overseas consumer-end AI application Talkie from MiniMax is one of its main sources of revenue. However, AI overseas applications still face regulatory policy challenges. Not long ago, Talkie was removed from the Apple Store in the USA, while the Android version has not been affected yet.

Compared to startups, large Internet companies have the advantage of social platforms, search engines, and other traffic entry points, allowing them to integrate with their own applications and be the first to achieve the landing and commercial realization of large models.

According to disclosures, the large model technology capabilities of ByteDance's Doubao support over 50 consumer-end application scenarios, including Doubao, Jimeng, and Doubao MarsCode. Through Volcano Engine, Doubao's large model serves over 30 industries, with an average daily tokens call volume exceeding 4 trillion, growing 33 times since its release in May.

Currently, ByteDance is adopting a strategy that simultaneously focuses on both the consumer and business ends. "Large models should progress seamlessly in technology and application, consumer and business ends, and physical and virtual worlds," said Tan Dai, president of Volcano Engine under ByteDance, in an interview. "Entertainment has to process information, knowledge acquisition requires information processing, work needs information processing, and enterprise operation also involves information processing; both 2C and 2B progress together. The physical world serves people and interacts with them, including auditory, language, and visual elements. These tasks were not well executed before, but large models can perform them well, connecting the digital and physical worlds more easily than many past technologies."

In 2023, Baidu initiated the integration of the Wenxin large model with Baidu Search, Input Method, Baidu Wenku, and others. However, compared to ByteDance and Tencent, Baidu faces challenges from the decline of search engines and even the disruption of AI search. The Star Daily noted that companies like 360 Group, Douyin, and Xiaohongshu have successively launched AI search products. To address this, Baidu is also strengthening its AI search layout, including recently launching the Baidu 'AI Search' (Chat.baidu.com) entry on the homepage of Baidu Search on the web, promoting the revision and upgrade of Baidu Search.

To further bridge the gap with end users, the previous Baidu Wenxin Yiyan APP has been renamed to Wen Xiaoyan, positioning itself as Baidu's 'new search' intelligent assistant, equipped with rich media search, multimodal input, text and image creation, and highly realistic digital human capabilities.

Among the domestic giants, Alibaba and Baidu are similar in that they started early on their large model strategy. However, they have also been surpassed by latecomers like ByteDance in end-user applications. One reason may be that Alibaba's efforts in large models for end users are relatively scattered, with various departments acting independently and not able to form an effective joint force.

In April 2023, Alibaba Group's then Chairman of the Board of Directors and CEO, as well as CEO of Alibaba Cloud Intelligent Group, Zhang Yong, stated that all Alibaba products would be integrated with the 'Tongyi Qianwen' large model for comprehensive transformation.

However, after Zhang Yong resigned as CEO and Chairman of Alibaba Cloud on September 10, 2023, news soon emerged about the Taobao Tmall Group under Alibaba preparing to establish a large model team. In March 2024, the self-developed large model 'Taobao Xingchen' official website launched. At that time, the website not only provided detailed information about the large model but also set up 'Experience Now' and 'Login/Register' entries. However, the Star Daily reporter found that the website currently shows a state of construction.

Informed sources told the Star Daily reporter that Taotian's Xingchen large model is developed independently by the Taotian Group and does not use the foundational model of Tongyi Qianwen.

Alibaba's Intelligent Information Division also released a fully self-developed, billion-parameter Quark large model in November 2023, applied in general search, medical health, education, and workplace scenarios. Subsequently, the Quark App established itself as a one-stop AI service centered on AI search.

Alibaba has clearly seen its shortcomings in end-user applications, choosing to adjust in early December 2024 by spinning off the AI application 'Tongyi' App from Alibaba Cloud and integrating it into the Alibaba Intelligent Information Division. This allows Quark and Tongyi Apps to develop as two parallel brands within the Intelligent Information Division. However, the Tongyi Laboratory and the Tongyi large model remain unchanged. This means that the underlying technology and product application teams of 'Tongyi' will be separated, with Wu Jia, President of the Alibaba Intelligent Information Division, taking on the responsibility for AI applications at end user level.

Looking towards 2025, the domestic competition among large models will continue. From the "battle of hundreds of models" to the current "application battle", the market is more eager for practical applications to emerge from the models, enabling large model vendors to earn profits and improving efficiency across various industries.

Baidu founder Robin Li also predicted in a letter to all employees on January 1, 2025, that "although super applications have not yet emerged, the actual penetration rate of AI is already considerable, and it will continue to explode in growth in 2025."

Editor/lambor