Text/Wang Huiying

Editor/Ziye

One evening in October 2016, at the Mercedes-Benz Arena in Shanghai, Luo Yonghao was busy launching the Smartisan phone. Unexpectedly, this launch event made iFLYTEK Input Method deeply rooted in people's minds.

Overnight, iFLYTEK Input Method surged to third place on the App Store's tool list and became the highest-ranked third-party input method. iFLYTEK, the company behind iFLYTEK Input Method, also gained popularity.

iFLYTEK is not a young startup, nor is its founder Liu Qingfeng an 'unknown'.

In 1990, Liu Qingfeng gave up his guaranteed admission to the Department of Automotive Engineering at Tsinghua University and instead enrolled in the Department of Electronic Engineering at the University of Science and Technology of China with a score 40 points above Tsinghua's admission line;

In 1999, the 26-year-old Liu Qingfeng turned down overseas doctoral opportunities with a $10,000 scholarship and a Microsoft scholarship to found iFLYTEK.

Liu Qingfeng and early iFLYTEK employees, image source: iFLYTEK WeChat official account

Over the past 20+ years, from an 18-person startup team to a listed company, iFLYTEK has become one of the largest AI technology vendors in China and a benchmark for domestic technology companies.

As AI's role in the industry continues to evolve, Liu Qingfeng has guided iFLYTEK through a long journey. This journey has had its highs and lows, competitors and allies, doubts and scrutiny.

In the second year of iFLYTEK's establishment, IDG Capital partner Lin Dongliang intended to invest in iFLYTEK, but Chairman Xiong Xiaoge left a judgment that "it looks lively but doesn't make money," and the cooperation fell through.

However, ultimately, iFLYTEK proved itself and became a leading Chinese AI company. Liu Qingfeng mentioned that iFLYTEK has taken a "curved straight line": curved refers to the specific path being very tortuous; straight refers to clear and definite goals and strategies.

At the end of 2022, the emergence of ChatGPT led to the explosion of generative AI, prompting Liu Qingfeng to decide to go all in.

Over the past year, from iFLYTEK Spark 1.0 to iFLYTEK Spark 4.0 Turbo, from technology to multimodal interaction to landing scenarios, iFLYTEK has surged ahead, becoming a pioneer in the domestic "Hundred Models War".

Especially this year, the commercialization of large models is accelerating, and the industry is entering deeper waters. As the elimination race intensifies, iFLYTEK's investments are paying off. In the third quarter of this year, iFLYTEK's revenue increased year-on-year, and its net profit turned positive.

'Light the brightest torch, and either be the first to set the prairie ablaze or the first to go out.' Liu Qingfeng said this at iFLYTEK's crucial 'Bantang Conference' in 2001.

In recent years, Liu Qingfeng has had new thoughts: 'Only with an autonomously controllable ecosystem can there be an ever-lasting future.'

The driving force behind Liu Qingfeng's years of adherence to technology is his love and original aspiration for the AI industry. The result of his years of adherence to technology is the confidence he has in leading iFLYTEK into complex, deeper waters.

1. With large model applications landing, Liu Qingfeng aims to take iFLYTEK into 'deeper waters'

On the day of the company's 14th-anniversary celebration, Liu Qingfeng invited all employees to watch 'American Dreams in China'.

He hopes that these young people, with an average age of less than 30 and just like him when he first started his business, will see two things from the film: foresight and perseverance.

Just like when he insisted on doing voice technology back then.

In 2001, iFLYTEK, which had been established for less than two years, was struggling, while real estate, gaming, and other industries were booming. In contrast, voice technology had a high threshold and was difficult to monetize, so some suggested that iFLYTEK should also layout related industries.

At that year's 'Bantang Conference,' Liu Qingfeng bluntly stated: 'iFLYTEK should do what iFLYTEK should do. If you don't like voice, please leave.'

2001 iFLYTEK Annual Plan Summary Meeting, image source: iFLYTEK WeChat official account

It was also at this conference that iFLYTEK determined its strategic direction – not only to excel in smart voice and AI but also to represent China in global competition.

This path has not been easy. iFLYTEK's layout has not been smooth, and its revenue has been around 10 billion yuan.

Remarkably, iFLYTEK's determination to continue on this AI path has not wavered.

In 2023, with the rise of large models sparked by ChatGPT, the opportunity for iFLYTEK has arrived.

At that time, Liu Qingfeng borrowed an account to experience ChatGPT and immediately decided to go all in. iFLYTEK's research institute mobilized its strongest team to research the iFLYTEK version of the large model while simultaneously promoting application landings.

Since its first release on May 6, 2022, the iFLYTEK Spark large model has undergone multiple iterations. By October 24, 2023, at iFLYTEK's Global 1024 Developer Festival, iFLYTEK Spark had been updated to version 4.0 Turbo.

According to iFLYTEK, based on real data back-to-back testing, the seven major capabilities of iFLYTEK Spark 4.0 Turbo comprehensively surpass those of GPT-4 Turbo, with mathematical and coding abilities exceeding GPT-4o, achieving first place in 9 of 14 mainstream test sets in Chinese and English domestically and internationally.

Image source: iFLYTEK WeChat official account

From a global industry perspective, after nearly two years of the 'Hundred Models War,' capital has entered a cooling-off period, and the industry has entered deeper waters. Pure technical parameters and vague concepts are no longer sufficient to support the large model industry in entering the next stage, and large-scale application landings have become the key to the new round of competition among players.

In other words, no matter how high the technical threshold and computing power level are, they must meet the basic demand of large-scale landing of large models.

'The exploration of future AI technology must be industrialized on a large scale and must enter true deeper waters in scenarios. All those who play with 'concepts' will not have a great future,' Liu Qingfeng said in his speech at this year's 1024 iFLYTEK Global Developer Festival.

On October 24, 2023, iFLYTEK launched 10 hardcore products and innovative applications based on the capabilities of the iFLYTEK Spark platform. It also showcased over 200 large model application products at the Science and Technology Expo, covering hundreds of scenarios such as education, healthcare, automobiles, finance, industry, enterprises, and personal assistants.

For example, in the office field, following the iFLYTEK translator, iFLYTEK officially launched a multi-language AI translation transparent screen; in the industrial field, the VIAS intelligent cockpit human-computer interaction effect evaluation robot can use AI to assist in vehicle inspection; in the education field, the Alpha Egg Go AI can more flexibly respond to children's personalized reactions in Go education...

Landing is not an overnight process. As a domestic large model vendor, iFLYTEK's technological first-mover advantage is undeniable. However, as the industry enters deeper waters of scenario-based landings, to gain an advantage in this battle, it is necessary to return to commercial landings and the essence of the industry.

2. To reach this point, how many hurdles has Liu Qingfeng led iFLYTEK over?

Throughout the journey, Liu Qingfeng has been well aware of the fierce competition faced by his industry and has enjoyed the feeling of competition.

Over 20 years ago, when doing voice technology, his competitors were companies like Microsoft and Google. Now that his former competitors have grown into giants, Liu Qingfeng has welcomed new competitors.

At the end of 2022, the emergence of ChatGPT gave Liu Qingfeng a clearer goal – to fully benchmark ChatGPT but to create an autonomously controllable model for iFLYTEK.

In this 'Hundred Models War,' domestic technology companies are all participating with their own products. It is no longer difficult to imitate ChatGPT, and the focus of this competition is who can create a better 'Chinese large model.'

Liu Qingfeng's thinking is clear: technically benchmarking, iFLYTEK provided the most detailed timeline for technological progress at its first large model launch event and fulfilled it as scheduled.

On May 6, 2022, iFLYTEK released version 1.0 of the Spark large model. At that time, Liu Qingfeng stated that the Spark large model would benchmark ChatGPT (GPT 3.5) by October.

Liu Qingfeng provided a detailed timeline, estimating that there would be three upgrades within 2023, enhancing mathematical capabilities in batches, supplementing coding capabilities, and aiming to surpass not only in Chinese but also reach the same level as ChatGPT in English.

Five months later, iFLYTEK released the 'iFLYTEK Spark' cognitive large model V3.0, and Liu Qingfeng excitedly stated, 'We fulfilled the timeline we promised everyone on schedule.'

By the first half of this year, iFLYTEK was benchmarking GPT-4. 'Within six months, we can reach the current best level of GPT-4 worldwide,' Liu Qingfeng set a new goal.

On June 27, iFLYTEK released the Spark large model V4.0, which significantly enhanced the model's capabilities and fully benchmarked GPT-4 Turbo in terms of its underlying capabilities.

Not long ago, on October 24, Liu Qingfeng again updated iFLYTEK's report card.

Behind each upgrade are various parameter comparisons, but other vendors are different. From the beginning, Liu Qingfeng believed that the landing application of large models was more important. Even in the early stages of the Hundred Models War, in addition to listing parameters, iFLYTEK also focused on the application of large models in its own products.

For example, in learning scenarios, multiple AI learning machines under iFLYTEK also incorporate large models, enhancing capabilities such as real-scene dialogue and advanced writing correction; in office scenarios, traditional speech-to-text services are further upgraded, utilizing the large model's natural language processing capabilities to enable automatic meeting minutes, text regularization, and one-click document generation.

The problem facing large model technology companies is that it takes time and cost to train products to move from the laboratory to market-scale production. It is unknown how many users will ultimately pay for this, but real investment cannot stop.

This undoubtedly puts pressure on iFLYTEK. Over the past year, due to continuous investment in large models, iFLYTEK has still incurred losses.

In the first half of this year, iFLYTEK achieved operating revenue of 9.325 billion yuan, a year-on-year increase of 18.91%. The net profit attributable to shareholders of the listed company was a loss of 401 million yuan, compared to a net profit of 73.572 million yuan in the same period last year, resulting in a switch from profit to loss.

iFLYTEK stated in its announcement that the loss was due to additional investments of over 650 million yuan in the first half of the year for large model research and development, core technology autonomy and control, industrial chain control, and industrial landing expansion. At the same time, investment income, fair value, and other income decreased year-on-year, and provision for bad debts increased year-on-year. The announcement showed that iFLYTEK's R&D investment in the first half of the year was 2.19 billion yuan, an increase of 32.23% year-on-year.

In Liu Qingfeng's view, increasing R&D investment stems from confidence in the future of large models. 'If we choose to be conservative and save on these investments, it seems like profits will immediately appear. We are not unaware; we are all good at math. After so many years of entrepreneurship, we are all shareholders of the company. Why not do it this way? The reason why we explore in many places is that we don't want to miss this major revolutionary opportunity and stand on the same platform as all top players,' Liu Qingfeng said.

In the third quarter of this year, Liu Qingfeng's confidence in the future began to shine through. iFLYTEK's third-quarter financial report showed that in the third quarter of 2024, operating revenue was 5.525 billion yuan, a year-on-year increase of 15.77%; the net profit attributable to shareholders of the listed company was 56.9614 million yuan, a year-on-year increase of 120.87%, with a positive net profit for a single quarter.

iFLYTEK's progress to this point is inseparable from Liu Qingfeng's thinking. When facing the gap with competitors, Liu Qingfeng will 'catch up; we know how to do it.' When facing questions about its technology investment from the market, Liu Qingfeng said, 'After understanding the temperament of large models over the past six months, we are now the most confident.'

Since the rise of large models, Liu Qingfeng, who has changed from his previous low-key role as a behind-the-scenes operator, has been present at every iFLYTEK large model launch event. This 'model student' in the AI world is again throwing himself into the fray, just like when he was young and working on voice technology.

3. With five keywords, where does Liu Qingfeng aim to take iFLYTEK?

If we dissect the entrepreneurial journey of Liu Qingfeng and other science and engineering prodigies, we will find that they are daring and ambitious individuals.

Especially in the era of entrepreneurial fever, outstanding students who excel in school are often not satisfied. Prodigies often have a 'Bill Gates' or 'Steve Jobs' living in their hearts.

The time goes back to 1992. Liu Qingfeng, who excelled in mathematics and physics, was still a sophomore when he was selected by Professor Wang Renhua, his mentor, to join the Human-Computer Speech Communication Laboratory at the University of Science and Technology of China. Here, robots could speak, and Liu Qingfeng planted the seeds of his ideal in the voice industry.

To this day, when talking about how iFLYTEK's voice technology has reached the top nationally and even globally, Liu Qingfeng still attributes it to his trust in the voice and AI industries back then.

Liu Qingfeng bluntly stated that this is entrepreneurship. 'Because we believe, we see,' rather than 'because we see, we believe.' There is a fire in our hearts, believing in the future of the industry, so we can see the path and the hope for the future.

The fire in Liu Qingfeng's heart has never gone out. At the opening ceremony of iFLYTEK's Global 1024 Developer Festival, Liu Qingfeng emphasized at the end of his speech that the key to the future development of the AI industry lies in five keywords: standing tall, autonomy and control, integration of general and specialized, end-cloud linkage, and software-hardware integration.

Image source: iFLYTEK WeChat official account

Breaking down these five keywords reveals more of Liu Qingfeng's thoughts on the landing of the AI industry.

The so-called 'standing tall' refers to innovating from source technology, aiming to be globally leading and benchmark the best underlying models and specialized models worldwide to achieve global first place;

'Landing' must involve large-scale industrialization and must enter true deeper waters in scenarios;

'Autonomy and control' mean that all those who play with concepts will not have a great future; only autonomy and control lead to a great future;

'Integration of general and specialized' refers to combining general models with specialized models to make them more efficient and effective in specialized fields;

'End-cloud linkage and software-hardware integration' truly unleash productivity and imagination.

Behind these five keywords lies the persistence of a pragmatist.

First, technically, iFLYTEK has its own adherence.

Over the past two years, there have been two important milestones in the large model industry: the emergence of ChatGPT and the emergence of Sora.

When the former appeared, iFLYTEK set a 'military order' to fully benchmark it. When Sora introduced video generation capabilities, Liu Qingfeng was not in a hurry to follow suit. Liu Qingfeng believes that for the multimodal capabilities of large models, generating videos is merely entertainment, and the higher priority is entering industrial scenarios, living scenarios, family companion robots, and other embodied intelligence.

In an interview with China Entrepreneur, Liu Qingfeng said, 'The multimodal understanding ability of large models is more important than multimodal generation. We also have similar plans to launch Sora, but this plan will come after the general large model.'

From the perspective of the core underlying technology of large models, 'pretrained models,' entering 2024, due to various restrictions, many large model companies have been rumored to abandon pretrained models. iFLYTEK is one of the few vendors that persist in making pretrained models.

Pretrained models are a technical task that requires high talent, technology, and cost. Like Li Kaifu, CEO of Innovation Works, mentioned, pretrained models require people who understand chips, reasoning, infrastructure, models, and algorithms to work together.

This also means that not every company has the ability to do this. iFLYTEK, which began working on pretrained models in 2019, is still iterating continuously. Liu Cong, Dean of the iFLYTEK Research Institute, said in an interview with 36Kr that iFLYTEK's large language model will continue to iterate, with continuous layout in areas such as o1 large models, multimodal interaction, and end-to-end speech.

Previously, Liu Qingfeng has emphasized on multiple occasions that, against the backdrop of the Sino-US technological and AI competition, the capabilities and support of domestic large model foundations determine whether one can enjoy more dividends from this wave of AI.

Under this mindset, iFLYTEK Spark is the first and only large model trained on a domestic computing power foundation. At this developer conference, iFLYTEK also launched the 'Feixing II,' a domestically produced super-large-scale intelligent computing platform jointly created by Huawei and Hefei Big Data Asset Operation Co., Ltd.

iFLYTEK said that last year's 'FeiXing No.1' resolved over 500 basic software and hardware issues and model adaptation problems, and this year's 'FeiXing No.2' will bring continuous adaptation of new models and algorithms and another leap in intelligent computing cluster scale.

Secondly, in terms of large-scale deployment, like many other large model vendors, iFLYTEK is pursuing development in both B-end and C-end markets.

At the B-end level, as of now, over 415,000 enterprise customers have utilized iFLYTEK Spark innovation applications. "iFLYTEK Spark has become the first choice for large models in central and state-owned enterprises," said Liu Qingfeng.

According to Liu Qingfeng, iFLYTEK has a leading toolchain that significantly improves the efficiency of "processing data and training models"; it also has practical application cases covering over 300 industry scenarios, forming a scalable effect of mutual learning and reuse.

At the C-end level, empowered by the iFLYTEK Spark large model, C-end hardware sales have increased. In the first three quarters of this year, iFLYTEK's sales of study machines, consumer and developer platform business, and smart car business increased by over 100%, 44%, and 49% respectively, effectively offsetting the impact of declining revenue from G-end business such as smart city information engineering.

In terms of the overall ecosystem, in the first three quarters of 2024, the number of new developers for the iFLYTEK Spark large model reached 1.777 million. According to data from QiMai, the iFLYTEK Spark App has been downloaded over 190 million times on Android.

In 2019, when iFLYTEK celebrated its 20th anniversary, Liu Qingfeng outlined his vision for the future of artificial intelligence in the next decade in his letter to all iFLYTEK employees, pointing out three major trends: voice will become the most important human-computer interaction method, the Tower of Babel for human language will be ultimately constructed within ten years, and artificial intelligence will empower various industries.

Today, Liu Qingfeng's vision remains unchanged, and the emergence of ChatGPT has given him even more confidence in the future of the AI industry. He aims to lead iFLYTEK not only in technology and scenario implementation but also in using AI technology to fulfill people's aspirations for the future.

However, achieving these goals requires a more combat-ready iFLYTEK. With two years left to reach iFLYTEK's goal of "one billion users, achieving ten billion in revenue, and driving a trillion-dollar industrial ecosystem" in the next five years, entering the "deep water zone," Liu Qingfeng is destined to face numerous challenges and overcome them one by one.

(The featured image of this article is sourced from the iFLYTEK WeChat official account.)