How long does it take to benchmark and surpass the so-called world's most advanced large models, GPT-4 Turbo/GPT-4o?

iFLYTEK's answer is 560 days. That is, from December 15, 2022, when iFLYTEK officially launched the "1+N" cognitive large model special research project, to June 27, 2024, when iFLYTEK Spark V4.0 was officially released.

Such an achievement is enough to attract attention from the outside world. It's worth noting that iFLYTEK achieved full-stack independent research and development, complete autonomy and control, and comprehensive localization in the face of extreme pressure from the United States, which is truly impressive.

Not only has the technology reached the world's first tier, but the company's commercialization landscape has gradually unfolded since its shift to general AI. All of this can be glimpsed from its growth in 2024.

【Quality Growth】

General AI has entered an era of contention and competition.

AIGC has become a hot spot pursued by global financial markets, but amidst the fervor, investors are placing more emphasis on the industrialization and actual performance of AI enterprises.

Due to the enormous computing power required by large models and the shortage of AI chips leading to soaring prices, even OpenAI has to bear huge losses.

Before the outbreak of general AI, Jen-Hsun Huang defied public opinion and insisted on continuous investment in CUDA, inviting Lan Buck to join NVIDIA, ultimately achieving today's glory as the "King of AI Chips".

Therefore, the profits of related enterprises should be viewed objectively and comprehensively. On July 6, iFLYTEK released its 2024 semi-annual performance forecast, which reveals the continuous progress of iFLYTEK Spark's commercialization.

The company expects revenue for the first half of the year to exceed 9 billion yuan, with a year-on-year growth rate of 15%-20%. Gross profit is expected to increase by 600 million yuan, up 15%-20% year-on-year, and the growth rate of gross profit is higher than that of revenue, indicating an improvement in real profitability, i.e., gross margin.

During the investment period, financial data such as revenue, gross profit, and collections are, to some extent, more important than profit. In the first half of the year, the company's sales collections amounted to approximately 9 billion yuan, an increase of 1.49 billion yuan year-on-year, further improving its cash flow situation.

Moreover, the business structure has been further improved, and financial health has significantly increased. C-end revenue has grown significantly. Since the release of the Spark large model, AI-empowered products such as iFLYTEK learning machines, smart office notebooks, and translators have experienced high-quality growth.

In the first five months of this year, iFLYTEK learning machine sales surged by over 150%. During the "618" e-commerce promotion, iFLYTEK smart hardware sales increased by 70% year-on-year, with an average monthly usage exceeding 40 million times.

It can be seen that iFLYTEK's focus on sustainable C-end operations has achieved remarkable results. The expansion of C-end business has gradually reversed the company's past reliance on G-end business, which is of great significance for improving the company's financial quality.

Generally speaking, G-end business often has longer account receivable cycles and faces many uncertainties such as policy changes and project delays. In contrast, C-end business is essentially a "cash on delivery" model, significantly improving the financial situation.

In the first half of this year, the company continued to invest heavily in R&D, coupled with the shrinkage of financial assets, investment income, and the provision for bad debts, resulting in a decline in profits.

However, on the other hand, it should be noted that high-intensity investment is necessary to maintain AI technology's autonomy, controllability, and leading position, as technology giants worldwide are investing. In this industry that determines future destiny, only continuous investment can ensure one's place at the table.

At the same time, the company also stated that in the coming years, the growth rate of additional investments in computing power for Spark large model research and development will be relatively small, and the pressure on operating fundamentals is manageable.

From intelligent voice to cognitive intelligence and then to general AI, iFLYTEK is in a critical period of strategic transformation. As Liu Qingfeng said, "The current state of general AI in China is very similar to iFLYTEK's start-up situation 25 years ago and the early development of China's voice industry."

Facing the vast sea of AIGC opportunities and combined with the improvement of business structure, iFLYTEK is gradually entering a track of high-quality development.

【Determination to Break Through】

AI has become the commanding height of today's and future global technology industries. As a leader in China's AI technology, iFLYTEK shoulders heavy responsibilities but has also suffered from stringent technical blockades.

"True gold comes from the crucible of trials." iFLYTEK was included in the US Entity List twice in 2019 and 2022, but this did not hinder its progress. Instead, it honed its ability to cope with extreme pressure.

The company adheres to technological independence and has joined forces with partners such as Huawei to build a national team of computing power. Together with Huawei Ascend, iFLYTEK established China's first domestic tera-card computing power platform, "Feixing No.1," and iFLYTEK Spark is currently the only large model trained on this domestic computing power platform.

"Without a general large model foundation, focusing only on industry models will erase existing advantages when a new generation of general large models is released." iFLYTEK has a clear and profound understanding of this.

As the US continues to tighten export controls on high-end AI chips and strictly restricts the sale of high-performance semiconductors to China by companies such as NVIDIA and AMD, whether iFLYTEK has the capability of an autonomous and controllable domestic foundation determines how far and steadily it can go on the AI path.

On December 15, 2022, iFLYTEK officially launched the "1+N" cognitive large model special research project, where "1" refers to the general cognitive intelligence large model, and "N" refers to the application of the large model in various fields such as education, office, automotive, and human-computer interaction.

With the support of "Feixing No.1," on May 6, 2023, Spark large model V1.0 debuted, marking the first step into the era of general AI. By June 27, 2024, iFLYTEK Spark V4.0 was officially released, fully benchmarking GPT-4 Turbo in terms of base capabilities. In 12 mainstream tests of large models in Chinese and English at home and abroad, iFLYTEK Spark ranked first in 8 tests, surpassing GPT-4 Turbo.

iFLYTEK Spark V4.0's image and text recognition capabilities have reached new heights, outperforming OpenAI's latest GPT-4o in applications in scientific research, finance, healthcare, judiciary, and office scenarios.

During this year's college entrance examination, a horizontal evaluation of large model capabilities was conducted using college entrance examination essay and math questions. iFLYTEK Spark not only ranked first but also surpassed GPT-4o.

It's worth noting that this was an almost "impossible mission" accomplished by iFLYTEK under US trade sanctions.

In just 560 days, iFLYTEK demonstrated China's technological prowess to the world.

Currently, the company is sparing no effort to increase investment and consolidate its leading achievements. In the first half of the year, it invested over 650 million yuan in large model research and development, core technology autonomy and controllability, industrial chain controllability, and large model implementation and expansion.

Since its full opening in September 2023, the "iFLYTEK Spark" App has ranked among the top despite investing far less than first-tier vendors. The number of downloads on the Android side has exceeded 142 million, ranking first among domestic tool-type general large model Apps (data from Seven Mac).

In terms of positioning and functionality, iFLYTEK Spark differs from other internet vendor large models that are entertainment-oriented, with more productivity attributes, growth potential, and commercial value.

As seen from user personas, iFLYTEK Spark users are primarily knowledge-based workers, with usage scenarios mostly being rigid demand scenarios. Usage frequency is concentrated during working hours, primarily for solving work-related problems, such as "copywriting rewriting, creative generation, and code programming." This better reflects the professional value of iFLYTEK Spark.

"Past experience is a teacher for future actions." ZTE and Huawei once suffered from technology blockades that hindered their development. The sustainable development of general AI also needs to be built on full-stack autonomy and controllability.

After 560 days of catching up, iFLYTEK firmly believes in technological self-reliance and has always maintained a leading position in the world. With its unique autonomy and controllability advantages, the enormous value hidden beneath the iceberg is just beginning to emerge.

【Unique Industrialization Path】

Since its inception, iFLYTEK has adhered to the technical concept of "reaching for the sky and standing firmly on the ground." "Reaching for the sky" refers to maintaining international leadership in core technologies, while "standing firmly on the ground" refers to realizing large-scale industrialization of technological achievements.

Currently, the company is accelerating the implementation of general AI technology and deepening its strategic layout of "platform + track" to accelerate the development of the AI ecosystem.

The company implements a "1+N" layout strategy, with "1" as the base, i.e., the general cognitive intelligence large model foundation, and "N" as the tentacles, applying cognitive intelligence large model technology in various vertical fields such as education, healthcare, automotive, office, finance, industry, robotics, and smart hardware, forming industry-specific models with unique advantages.

Data is the most powerful proof.

In the second quarter of this year, iFLYTEK demonstrated strong competitiveness in industries such as central and state-owned enterprises, operators, and finance, winning major contracts exceeding 10 million yuan from clients such as Guoyuan Securities and China Mobile Sales Company.

▲Image source: Intelligent Hyperparameters

In the first half of this year, iFLYTEK showed a fierce momentum. According to incomplete statistics, the company won at least 14 large model projects. Especially in June, it won 8 projects with a total bid amount of 38.95 million yuan.

Thanks to its full-stack autonomy and controllability, central and state-owned enterprises and large state-owned banks favor iFLYTEK. Among the 14 bidding enterprises, at least 7 are from central and state-owned enterprises.

Benefiting from its previous accumulation in vertical industries, industry precipitation, and customer relationships, iFLYTEK has also secured orders from industries such as justice, operators, and finance, which are sectors where the company has long been deeply engaged.

In terms of implementation, iFLYTEK has launched multiple vertical models in niche areas based on its computing power foundation. The problem with general large models is that they primarily learn and train on general, publicly available knowledge, making them somewhat strained when dealing with vertical industries with information barriers.

iFLYTEK addresses this by adopting a "specialized, segmented, and vertical" approach to clear the path for large model implementation.

Focusing on both C-end and B-end, iFLYTEK launched "Personal Space" for C-end users, bridging its entire ecosystem of C-end software and hardware products, enabling users to access the "Spark Suite" with a single click.

In response to the increasing demand for health, iFLYTEK upgraded its iFLYTEK XiaoYi App, released a personal digital health space, and created an AI health assistant for everyone and every family, with downloads exceeding 12 million times.

On the B-end, iFLYTEK is committed to solving the last-mile problem of "AI+" for enterprises by launching the Spark Enterprise Intelligent Agent Platform, which creates dedicated assistants for each position using the intelligent agent platform.

The intelligent agent platform can quickly build implementable applications based on business scenarios and has launched 32 enterprise intelligent agents covering production, scientific innovation, office, and management scenarios, cutting into the fine details of enterprise business operations and solving pain points. This allows enterprises to plug and play, reducing the cost of integration.

Typical applications include Spark Business Opportunity Assistant and Spark Bid Evaluation Assistant.

For example, applications have been effective in multiple typical scenarios such as code, compliance review, customer service, bid evaluation, and intelligent interaction. iFlyCode, a product based on Spark large model capabilities, covers more than 6,000 R&D personnel at Bank of Communications, with a code adoption rate of 38%, significantly improving work efficiency.

Based on the Enterprise Intelligent Agent Platform, Spark Business Opportunity Assistant can help enterprises discover business opportunities, improve customer visit efficiency, and provide intelligent analysis for sales management, among other functions. Spark Bid Evaluation Assistant achieves a 98% consistency rate between human and machine bid evaluation results, with an abnormal bid detection rate exceeding 80%, significantly improving enterprise bid evaluation efficiency.

Speech recognition is inherently iFLYTEK's strength. In the era of the Internet of Everything, Spark Speech Large Model has launched multi-language and multi-dialect speech recognition capabilities without switching, supporting free conversation in 37 languages and 37 dialects, achieving an 86% recognition accuracy rate even in scenarios with three people speaking simultaneously.

Such intelligent large models are driving changes in human-computer interaction, and all applications in the speech field will be reconstructed. With the support of large models, Spark Automotive Intelligent Cockpit has been upgraded, offering multi-language and multi-dialect "free interaction" as well as multi-emotional and multi-modal hyper-anthropomorphic interaction, making human-vehicle interaction more intimate. Relying on this technological advantage, iFLYTEK's speech interaction products have consistently ranked first in domestic market share and have empowered automakers such as FAW, Chery, GAC, JAC, and Great Wall with highly intelligent interaction experiences, helping them win in the second half of the intelligent race.

At the same time, iFLYTEK attaches great importance to the prosperity of the developer ecosystem. In the past five months, the number of iFLYTEK Spark developers has grown from 5.98 million to 7.02 million, an increase of over 1.04 million; the number of overseas developers exceeds 400,000; and the number of large model developers reaches 570,000.

"Only an autonomously controllable and prosperous ecosystem can lead to a bright future for China's general AI." As the Spark ecosystem flourishes, more application values in rigid demand scenarios will be unleashed.

Currently, general AI is still in the "Yahoo moment" of the Internet era, and the "rise of Google" and "domination of Tencent" are far from coming.

With both industry accumulation and technological advantages, as well as a unique methodology for industrial implementation, iFLYTEK is one of the most competitive enterprises at the AIGC table.

Disclaimer

The content related to listed companies in this article is based on the author's personal analysis and judgment of information (including but not limited to interim announcements, periodic reports, and official interactive platforms) disclosed by listed companies in accordance with their legal obligations. The information or opinions herein do not constitute any investment or other business advice, and Market Value Observer shall not be held responsible for any actions taken based on this article.