In the era of general artificial intelligence, the intelligence exhibited by large models carries significant practical significance and influence, marking a milestone in the development of artificial intelligence. With massive parameters, complex algorithms, and robust training capabilities, the competition among large models is heating up.
Recently, iFLYTEK released its 2023 annual report, achieving a total operating income of 19.65 billion RMB, a year-on-year increase of 4.41%. The year's fourth quarter saw a noticeable rebound with a 14% year-on-year growth. The net profit attributable to shareholders reached 660 million RMB, representing a 17.12% increase year-on-year. In 2023, the company invested 3.84 billion RMB in research and development, marking a 14.36% increase year-on-year and accounting for 19.53% of the revenue, an increase of 1.7 percentage points from 2022.
Increased investment in research and development yielded immediate results: iFLYTEK upgraded and released the Spark large model four times during the year and collaborated with Huawei to create the first domestically produced computing power platform, "Feixing 1," supporting the training of trillion-parameter large models. By jointly building a new foundation for general artificial intelligence in China, iFLYTEK enabled domestic large model architectures to be built on a foundation of independent innovation in both software and hardware, achieving autonomous control over core large model technologies.
During the latest performance briefing, Liu Qingfeng, Chairman of iFLYTEK, stated that the company had formed a unique competitive advantage in building high-performance models based on domestic computing power.
Saturated Investment in Large Model Development
The "Government Work Report" at the 2024 National Two Sessions clearly stated the need to "give full play to the advantages of the new type of nationwide system, and support capable enterprises in leading major breakthrough tasks." As a key player in the artificial intelligence national team, iFLYTEK is expected to receive more support and attention.
Since 2015, iFLYTEK has advocated the "AI+" strategy and has since demonstrated leading advantages in various capability tests and product capabilities. The company has set clear benchmarks: it aims to demonstrate the best level comparable to GPT-4 by June of this year in three aspects:
During the Two Sessions this year, Liu Qingfeng, a National People's Congress delegate and Chairman of iFLYTEK, emphasized that in the current competition dominated by large models in general artificial intelligence, only by building an independently controllable domestic computing power foundation can one grasp their own destiny and narrow the international gap.
Liu Qingfeng has repeatedly emphasized the need to address the gap while accelerating catch-up efforts. iFLYTEK, listed early on in the entity list domestically and subject to extreme pressure as a Chinese technology company in 2022, has continued to invest in independent and controllable aspects for many years. It has faced many challenges in its domestication strategy and has developed a unique competitive advantage in building high-performance models based on domestic computing power.
Last year, iFLYTEK joined forces with Huawei to create the first domestically produced computing power platform, "Feixing 1." The Xunfei Xinghuo 3.5, released on January 30, 2024, was also the first hundred-billion-parameter large model trained on a domestically produced computing power platform, leading in various professional tests domestically.
During the performance briefing, the management once again expressed their determination to increase research and development efforts and stated that the company would maintain a high level of investment in Spark large model development in the coming years. The Spark large model continues to benchmark against the international first-class level in general capabilities, ensuring that iFLYTEK continuously strengthens its core advantages in key industries and segmented scenarios such as education, healthcare, and office applications, thereby promoting business growth and improving efficiency.
Based on domestic large models, the Spark large model is empowering various industries and entering countless households. In the long run, large models are bringing about disruptive changes in fields such as healthcare, finance, and scientific research, promoting high-quality development.
According to company management, since the launch of the "1+N" large model research and development special project at the end of 2022, internal efforts have been rapidly advancing through eight working groups, with 70% of research and development focused on enhancing foundational capabilities. In 2023, the ratio between basic large model research and industry application research was roughly 7:3.
By 2024, it is expected that 50% of research and development investment will be used to ensure that the foundational capabilities of large models always represent the best level in China, continuously benchmarking against peers in the United States. It is expected that by June of this year, the Xunfei Xinghuo will be able to benchmark against the current best level of GPT-4 Turbo. In the future, the proportion of investment around basic large model research may decrease to 30% in research and development, while investment in applications may increase. The scale of investment in basic large model research will continue to grow with the expansion of industrial applications.
Highly Praised by Users
In 2023, direct research and development investment related to Spark exceeded 2 billion RMB, while market investment was only one or two hundred million RMB. Despite relatively low market investment, Spark still received high praise: the Spark app led in both reviews and ratings on the Apple App Store and ranked second in cumulative downloads on Android (Xunfei Spark: 927.82 million, Wenzhen One Word: 87.2751 million, Zhipu: 49.7531 million, Tongyi Qianwen: 12.9654 million; source: Qimai Data).
With increased research and development investment, there was a direct impact on performance growth. With the support of the Spark large model, many of iFLYTEK's businesses saw significant growth in 2023. The revenue contribution from the open platform and consumer business continued to increase, reaching 6.19 billion RMB, a year-on-year increase of 33.4%, becoming the largest business segment of the company, accounting for 31.47% of total revenue, exceeding 30% for the first time.
In the first quarter of this year, revenue from iFLYTEK's consumer business grew by 63%, gross profit by 48%; automotive business grew by 71%, gross profit by 46%; medical business revenue grew by 110%, gross profit margin by 127%; and C-end hardware learning devices in the education business grew by 99%, with gross profit growing by 104%. All business sectors have good development prospects. With the evolution of technology, the ability of iFLYTEK to empower various industries continues to improve, with product capability being the key. According to the company's management during performance briefings, large models empower the company's existing software and hardware products, as well as industry systems, further enhancing the company's leading advantages in related fields, improving user experience and product penetration, and bringing greater market potential. In various business areas such as education, healthcare, platforms, consumer services, and telecommunications, the company's original products have far from reached the ceiling of market potential in these industries.
Secondly, general capabilities can be fully empowered and will be implemented in more and more essential areas. The Spark large model collaborates with leading enterprises in various industries such as automotive, finance, energy, software, home appliances, and telecommunications.
For example, the code capability of the Spark large model has been tested by companies like Huawei and the Bank of Communications and is leading in the industry. In the software industry, since the release of iFlyCode on August 15, it has been applied in more than 100 institutions including Huawei, Pactera Technology, and the Bank of Communications.
iFLYTEK has currently joined forces with leading companies in various industries such as Chery Automobile, Pacific Insurance, Bank of Communications, State Energy Group, and China Mobile to jointly create industry benchmarks for large model applications in various fields, accelerating industrial upgrading. The company will release a heavyweight update on April 26 focusing on the analysis of long texts, multiple documents, long speeches, summaries, and question answering capabilities.
Thirdly, nurturing more entrepreneurs and sharing profits in an appropriate manner. The company provides a full-stack, self-controlled, and free Spark optimization suite for developers, universities, and enterprises, which not only deeply adapts to domestic computing power but also leads in scenario application effects, helping academic cooperation and industrial exploration, and being the first to apply large models to successful products.
Fourthly, in 2022, iFLYTEK launched the Super Brain 2030 plan, hoping that robots can enter thousands of households, which is a huge application direction. The progress of cognitive intelligence large models currently verifies the prediction of technical evolution in multi-modal perception and multi-dimensional representation in iFLYTEK's Super Brain plan. Based on the full-stack robot capabilities of the Spark large model, it has empowered 372 humanoid robot customers.
According to industry forecasts, after experiencing the entrepreneurial boom period, the market for general large models faces severe homogenization competition, and resources, technology, talent, and business landing capabilities have become the core competitiveness for manufacturers to move to the next stage. It is estimated that by 2025-2026, the market will gradually be cleared, with leading companies holding their ground while remaining companies gradually withdraw from the competition. By 2027-2028, it is expected that the market for general large models will present an oligopoly competition pattern, with 3-5 manufacturers becoming the cornerstone of China's large model ecosystem.
