By Chen Feng

Edited by Ziye

After nearly two years of fierce competition, China's large model sector is undergoing rapid differentiation.

This differentiation is first evident in the diverging paths chosen by large model startups. In 2024, the once highly publicized "Six Little Tigers of Large Models" embarked on distinct trajectories.

BaiChuan Intelligence pivoted towards industry-specific large models; Moon's Dark Side and MiniMax prioritized consumer-end (C-end) products and applications; Zero-One Everything adopted a "Big Factory + Little Tiger" collaboration model, entrusting the training of ultra-large models to Alibaba while focusing on small-parameter, moderately-sized industry models; Zhipu AI and Stepwise Stars remained focused on AGI large models.

From this perspective, it becomes increasingly clear that the "Hundred Model War" in China's large model industry is rapidly giving way to a new era. From startups to large companies, nearly all players are contemplating their roles in the wave of large models and AI, how to best implement technology, whether to focus on the domestic market or expand overseas, and how to create differentiated value.

Whether it's targeting consumers or businesses, the industry has reached a new consensus for 2025: large models will continue to move beyond homogeneity, entering a stage where technical prowess, implementation progress, and commercialization compete head-on.

Based on 2024 market performance, ByteDance and iFLYTEK stand out as noteworthy examples for discussion.

ByteDance has made a striking impression in the C-end market over the past year, demonstrating a rapid ascent to prominence. Globally, in terms of monthly active users, Doubao App has become the second most popular AI application after OpenAI's ChatGPT.

On the other hand, iFLYTEK has taken the lead in the business-end (B-end) sector.

According to the "2024 Monitoring Report on China's Large Model Bidding Projects" released by the third-party institution Smart Hyperparameter, in 2024, iFLYTEK topped the general large model vendor bidding rankings with 91 winning projects and a disclosed winning amount of 847.808 million yuan, making it the bid king of 2024 – in terms of disclosed winning amount, it was twice that of Baidu and eight times that of Zhipu AI.

Image source: Smart Hyperparameter

As current leaders in China's large model industry for the C-end and B-end, respectively, their paths have not been easy, but they both offer valuable insights.

1. Leading the Way on Two Distinct Tracks: iFLYTEK and ByteDance

In the large model race, ByteDance may not have entered early but has developed rapidly.

According to incomplete statistics from the DataEye Research Institute, since August 2024, ByteDance has launched a total of 17 large models and 2 intelligent agent development platforms in the AI field, including the Doubao large model family.

More C-end AI applications under ByteDance are also accelerating their launch. Since 2024, ByteDance has rolled out over 20 apps, including Doubao, in China and abroad, covering AI chat assistants, AI video tools, AI entertainment apps, office applications, and more.

Image source: Zheshang Securities

This aligns with ByteDance's past strategy of "miracles through force." Zheshang Securities statistics show that ByteDance's capital expenditure on AI reached 80 billion yuan in 2024, nearly matching the combined total of Baidu, Alibaba, and Tencent (about 100 billion yuan).

Market research firm Omdia's research also indicates that ByteDance purchased approximately 230,000 NVIDIA chips in 2024, becoming the second-largest global buyer of NVIDIA chips after Microsoft.

With greater investment and broader deployment, ByteDance quickly caught up in C-end applications over the year.

As of November, Doubao App's monthly active users had reached nearly 60 million, with an MAU growth rate of 16.92%.

On the other hand, in the less visible B-end large model market, iFLYTEK has steadily accumulated a leading edge.

Unlike ByteDance's "late entry, early success," iFLYTEK, to some extent, belongs to the "early entry, early success" category after the large model wave hit.

Shortly after OpenAI released ChatGPT at the end of 2022, iFLYTEK swiftly followed up on its large model layout. Over the next two years, it swiftly determined its technical approach and roadmap, completing multiple rounds of technical iterations.

Just half a month after OpenAI released ChatGPT, iFLYTEK had already decided to allocate resources to developing large models. At the same time, it also proposed that developing large models should be "1+N," where "1" refers to a general cognitive intelligence large model, and "N" refers to implementing these models in various fields such as education, office work, automobiles, and human-computer interaction.

At that time, iFLYTEK quickly assembled teams from 15 directions on its core R&D platform, specifically establishing a large model special group, which was further divided into four project groups focusing on "computing power and training frameworks," "data construction," "inference frameworks and services," and "algorithm research and development and large model creation." Hu Guoping, Dean of the iFLYTEK Research Institute, later recalled, "Such a large-scale 'battle' is rare in the history of the iFLYTEK Research Institute."

After that, iFLYTEK's iFLYTEK Spark large model accelerated its iterations.

On January 15, iFLYTEK officially released the Spark Deep Reasoning Model x1, along with the Spark Simultaneous Interpretation Large Model. Additionally, the base capabilities and industry capabilities of iFLYTEK Spark 4.0 Turbo were upgraded once again.

Earlier, iFLYTEK released iFLYTEK Spark 4.0 Turbo, with its seven core capabilities comprehensively surpassing GPT-4 Turbo, its mathematical and coding abilities surpassing GPT-4, and achieving first place in nine of the 14 mainstream test sets in Chinese and English, both domestically and internationally.

In terms of large model implementation progress, iFLYTEK is also at the forefront of the industry.

On the one hand, as mentioned above, in 2024, iFLYTEK was the "bid king" among general large model vendors;

On the other hand, also in 2024, iFLYTEK's Spark large model achieved six "firsts": first in central and state-owned enterprise bids, first in education and healthcare markets, first in smart automobile markets, first in large model developer ecosystems, first in smart hardware markets, and first in enabling scientific research applications.

After two years of land grabs, the industry landscape has become clear on the two paths of C-end and B-end – ByteDance and iFLYTEK have taken the lead.

2. Behind the "Winning Bids," How Do iFLYTEK and ByteDance Solve Problems?

Objectively speaking, neither ByteDance's resurgence in large model C-end applications nor iFLYTEK's B-end implementation exploration have been easy.

For C-end large model applications, the first challenge is high inference costs. Then, when considering product market fit, vendors must consider factors such as technical requirements, technical difficulty, and cost, and also need to seize the time window. Additionally, when exploring commercialization paths, compared to foreign countries, domestic users' willingness to pay is relatively weaker.

In other words, creating a user-friendly large model application that users love means higher input costs and an uncertain return period. This is why many large model startups have turned their gaze overseas in the past two years.

How to continuously attract new users and increase user retention rates is another challenge.

Judging from ByteDance's active layout in the large model field, it clearly aims to become a pioneer in creating more popular applications with greater potential opportunities.

First, ByteDance doesn't lack funds, technology, talent, or the determination to invest;

Second, ByteDance's successful C-end experience in the early years of the mobile internet era has now become its differentiated advantage. For example, compared to competitors, ByteDance has more abundant traffic to support rapid application growth.

Finally, ByteDance is becoming more keenly aware of user needs and more agile in its response speed.

On December 11, according to media reports, ByteDance elevated the product priority of JiMeng, attempting to create a "TikTok" of the AI era with a new path – JiMeng AI belongs to ByteDance's Jianying business and is positioned as an AI content platform that supports the generation of high-quality images and videos through natural language and image inputs.

It is reported that ByteDance plans to shift more resources to product forms with more modalities in the future, with JiMeng bearing greater hopes.

Now let's turn to the B-end. Today's competition in large models has gradually evolved into a battle of systems – to build a large model that enterprises can truly use, it is necessary to possess the entire suite of capabilities, including computing power construction, data governance, model training, scenario implementation, application construction, continuous operation, and security compliance, as well as the ability to create various standardized software products, such as digital humans, customer service assistants, and code assistants, as well as hardware-software integrated products in scenario implementation.

In short, the difficulty of large model implementation in the B-end lies in "delivery," requiring large model vendors to first become "hexagonal warriors."

From an enterprise perspective, at this stage, everyone's demands for large models are becoming more pragmatic, not only focusing on the advanced nature of model technology but also on how to integrate it into business scenarios and how to reduce costs and increase efficiency to solve practical problems.

Image source: "2024 China Industry Large Model Market Report"

iFLYTEK's problem-solving approach provides us with an observation window into the implementation of large models in the B-end.

"Why are we number one in winning bids, and why is our winning bid ratio increasing? Because many enterprises can only achieve the third step, which is model training, and they lag far behind us in the subsequent steps. Even if they can achieve them, their actual abilities in data organization and model training are still far behind us," Liu Qingfeng, founder of iFLYTEK, previously stated.

This corresponds to iFLYTEK providing a complete solution from top-level planning to implementation for enterprise large model construction: "Building computing power, organizing data, training models, implementing scenarios, ensuring security, and refining operations."

At the computing power level, in 2023, iFLYTEK and Huawei jointly built China's first 10,000-card computing power cluster, "FeiXing No. 1," which overcame many difficult issues on the basis of Ascend 910B, solving over 500 basic software and hardware problems, model adaptation issues, etc., enabling large model training to be improved from 20%-30% compared to A100/A800 to over 90%.

In October 2024, the domestic ultra-large-scale intelligent computing platform "FeiXing No. 2," jointly built by iFLYTEK, Huawei, and Hefei Big Data Asset Operation Co., Ltd., was also officially launched, which will bring continuous adaptation of new models and algorithms, as well as another leap in intelligent computing cluster scale.

The newly released Deep Reasoning Model X1 is based on "FeiXing No. 1" to create a deep reasoning model training framework that is fully compatible with Huawei's Ascend computing power, breaking through technical challenges such as tree search acceleration and asynchronous inference scheduling, achieving industry-leading results with less computing power, and ranking first domestically in multiple indicators, marking another key milestone for domestic computing power clusters to benchmark NVIDIA clusters.

At the "data organization and model training" level, iFLYTEK's complete tool chain has also significantly improved efficiency – data cleaning efficiency has been increased by 24 times, data construction efficiency by 90%, average scenario optimization effect by 30%, and knowledge editing efficiency by 5 times.

Then, in the more critical implementation of industry scenarios, as of October 2024, iFLYTEK has jointly built over 20 industry-specific large models with leading enterprises, covering over 300 application scenarios.

Liu Qingfeng also mentioned that these already implemented practical application cases have formed a scale effect of mutual reference and reuse. "After each enterprise sets it up, we will find many reusable elements for other enterprises. Many leading central and state-owned enterprises, after completing their work in this industry, can promote it to the entire industry and draw lessons from different industries."

It can thus be foreseen that the implementation of iFLYTEK Spark large models in the B-end, to some extent, is like a spark. From a long-term perspective, it may bring more ample imagination space to iFLYTEK.

3. The Commercialization Exam Approaches, and Leaders Accelerate Towards a "Positive Cycle"

As the large model race progresses, another increasingly clear fact is that elimination rounds have gradually begun.

Against this background, whether it's the C-end market or the B-end market, in the increasingly intense market competition, there is basically only one path for leaders to continuously maintain their competitive advantage and for chasers to erase the gap and catch up:

Maintain keenness on the technical side, fight a "protracted war" in iteration and upgrade speed, and excel at integrating large model technology with applications and scenarios.

This is exactly what iFLYTEK and ByteDance are doing.

During the recent iFLYTEK Global 1024 Developer Festival, in addition to releasing iFLYTEK Spark 4.0 Turbo, iFLYTEK also debuted 10 products and innovative applications based on iFLYTEK Spark's base capabilities:

These include defining the multi-modal AIUI standard, releasing ultra-humanoid digital humans, launching the Spark multi-language large model, iFLYTEK Spark medical imaging large model, automotive side-mounted Spark large model, and more.

Then, on January 15, iFLYTEK ushered in several new technological upgrades.

The Spark Deep Reasoning Model X1 has been unveiled, marking it as the sole deep reasoning model currently available on the domestic computing power platform. X1 has demonstrated its prowess by participating in numerous exams, spanning from elementary to high school (including competitions), university (including competitions), AIME, and MATH500, achieving remarkable results. Despite using less computing power, it has secured industry-leading positions, ranking first domestically across multiple indicators.

Furthermore, iFLYTEK has introduced the Spark Simultaneous Interpretation Large Model, the nation's first large model with end-to-end simultaneous interpretation capabilities.

The Spark Simultaneous Interpretation Large Model supports various translation modes with differing delays. During a rigorous 5-hour audio and video test, the 8-second delay mode outperformed foreign mainstream large models like Google's Gemini 2.0 and OpenAI's GPT-4 in aspects such as content completeness and information accuracy.

Based on the specific requirements of industry leaders and real-world feedback from over 200 million C-end users, iFLYTEK Spark 4.0 Turbo has undergone another round of enhancements, resulting in comprehensive improvements across its seven core capabilities.

Among these, the upgraded iFLYTEK Spark 4.0 Turbo boasts a 3.2% increase in text generation capability, a 4.5% improvement in language understanding, a 4.7% boost in knowledge Q&A, a 2.6% enhancement in logical reasoning, a 10.5% jump in mathematical ability, a 3.5% rise in coding ability, and a 1.6% improvement in multi-modal ability.

Moreover, iFLYTEK Spark has been upgraded in its handling of long text and image-text capabilities. It has also introduced mixed-domain knowledge search technology, allowing for comprehensive search results with just one question, whether it pertains to personal knowledge, enterprise knowledge, business system data, high-quality industry data, or internet information. This significantly enhances information search efficiency.

It is evident that with the continuous upgrading of its foundational capabilities, iFLYTEK's reach in various industries and scenarios in the B-end market is still expanding, its depth is still increasing, and its value is continuously being unlocked.

In the medical industry, the intelligent medical imaging assistant, built upon the iFLYTEK Spark medical imaging large model, aids imaging technicians in swiftly assessing image quality and promptly rectifying issues during intelligent quality control. It assists imaging doctors in rapidly generating diagnostic reports during intelligent diagnosis and helps clinicians devise treatment plans through related Q&A during intelligent image reading.

In judicial scenarios, the legal large model empowers various judicial tasks such as court trial transcript preparation, judgment document compilation, and legal case retrieval. Compared to the Spark general large model, efficiency has soared from 61.7% to 87.9%.

Throughout this process, iFLYTEK has clarified its long-term vision for the implementation path of large models.

For instance, regarding the optimal implementation of large model capabilities, Liu Qingfeng previously stated, "Today marks the dawn of a new era in large model implementation, one that harmoniously integrates general and specialized models, edge and cloud models, and software and hardware."

Not only does iFLYTEK aim to be at the forefront of the large model wave but it also strives to accelerate the progress of more enterprises. The assistance provided by the first simultaneous interpretation large model to Chinese enterprises venturing abroad serves as a prime example of this commitment.

To some extent, this aligns with the "positive cycle" path that ByteDance is currently accelerating. Through larger-scale and more decisive investments, ByteDance propels the rapid iteration and upgrading of large model technology, thereby empowering both C-end users and B-end customers. Although the return cycle for C-end users may be relatively long, they trade patience for future rewards, utilizing commercialization to support technological investments.

Upon this technological foundation, the logic behind ByteDance's C-end applications and iFLYTEK's deep dive into the B-end market are essentially similar—iFLYTEK endeavors to be closer to its customers, while ByteDance strives to be closer to its users.

This approach to products and services is the driving force behind ByteDance's meteoric rise to prominence within just half a year and iFLYTEK's consistent leadership in the field, illustrating a philosophy of "one step ahead, all the way ahead."

(The featured image of this article is sourced from the ByteDance official website and iFLYTEK's official Weibo account.)