Review of China's AI Development in 2024: The Road to Catch Up and Surpass

在人工智能发展史上，2024年注定是一个值得铭记的年份。这一年，中国AI企业上演了一场惊心动魄的追赶与超越之战。从年初紧追GPT-4的步伐，到年中直面GPT-4o的冲击，再到年末与OpenAI的o1系列针锋相对，中国企业展现出了惊人的执行力和创新潜力。

This is the era of the rise of the heroes. The "AI Six Little Dragons", including Smart Spectrum AI, Dark Side of the Moon, MiniMax, Baichuan Intelligence, Zero-One Everything, and Step Star, have been competing fiercely with Internet giants such as BAT. The favor of capital, the flow of talent, and technological breakthroughs are intertwined, constituting the most exciting chapter of this industry. When the bell rang at the end of the year, Chinese AI companies had already realized a real transcendence of international giants on certain tracks.

In this silent race, every OpenAI innovation provokes a wave of catching up. However, Chinese companies are not simply imitating, but gradually coming out with their own characteristics. From the full-stack layout of Smart Spectrum to the extreme experience of Dark Side of the Moon, from MiniMax's pursuit of efficiency to DeepSeek's vertical breakthroughs, each enterprise is looking for a unique development path.

This is a story full of drama and a saga of innovation and persistence. Let's go into 2024 and relive this exciting competition that changed the landscape of China's AI industry...

The January Herd: A Race to Start the Year Off Right

The first snow of 2024 has not yet fallen in Beijing, and the competition in the AI industry is already white-hot. In this industry, people have long been accustomed to the rhythm of continuous work - after all, OpenAI never gives opponents a chance to breathe.

It has been 9 months since the release of GPT-4, and this monument in front of all enterprises is still daunting. Although Baidu declared in October 2023 that Wenxin Yiyin 4.0 was "no less impressive" than GPT-4, the market and users are clearly still waiting for more challengers to come on the scene.

January 16, this seemingly normal working day is destined to go down in the annals of China's AI development. Smart Spectrum AI, founded by Tang Jie, a professor at Tsinghua University's Department of Computer Science, launched on this day theNew Generation Base Large Model GLM-4. When the test data came out, the whole industry was excited: in many authoritative reviews such as MMLU, GSM8K, etc., GLM-4 reached the level of GPT-4 90% or above. What's more remarkable is that GLM-4 has realized the first time to surpass GPT-4 in Chinese alignment capability.

Coincidentally, MiniMax, known in the industry as one of the "Six Little Dragons of AI", chose to show its sword on the same day. They released ABAB6, China's first MoE (Mixed Expert Model) large language model, an innovative mixed expert architecture that shows unique advantages in handling complex tasks and improving computational efficiency.

Just two weeks later, MiniMax struck again, launching abab-speech-01, a speech macromodel capable of completing tone reproduction in 6 seconds, demonstrating their strength in the multimodal field with the results of millions of hours of audio data training.

Since then, voice replication has also become a standard feature of domestic AI chatbots.

On January 22, a heavyweight entered the fray with his new work. Zero One Everything, led by Kai-Fu Lee, released Yi-VL Multimodal Language Model, a product that, upon its launch, created great results on the MMMU and CMMMU datasets, demonstrated strong cross-disciplinary knowledge comprehension, and catapulted the low-profile company into the spotlight of the global open source community.

仅仅四天后，被誉为”最懂AI”的团队月之暗面展示了他们对产品打磨的极致追求。Kimi Chat的重大更新不仅全面提升了基础模型能力，更在用户体验上实现突破：回复速度翻倍，联网搜索、上下文学习、多场景能力一应俱全。特别是小程序版本新增的语音输入功能，让这个产品在用户友好度上遥遥领先。

The perfection of these innovative features has quickly made this ever-so-low-profile company a noteworthy newcomer to the industry. In the first half of 2024, Kimi became one of the most popular AI apps, attracting a large number of users who used to focus on ChatGPT to switch to it. Its influence was so great that it even set off a "Kimi concept stock" investment boom in the capital market.

On January 29th, as this magnificent January was coming to an end, a heavy news triggered the whole scene. Baichuan Intelligence, dubbed "China's most maverick AI unicorn", launched Baichuan 3, a behemoth with over 100 billion parameters, which lived up to its name by demonstrating the best performance in healthcare tests, outperforming the GPT-4 in Chinese language tasks, and also performing well in natural language processing and code generation. The

At this point, the first month of the new year's tug-of-war is gradually coming to an end. Five "AI six little dragons" have made their debut, each showing a unique technology route and product features, while another six little dragons continue to build up their strength at the moment, and he will make his debut only a few months later.

However, right in the middle of all the cheering, a thought-provoking voice came from another tech giant.

On January 30, Liang Rupo, CEO of ByteDance, rarely revealed a strong sense of crisis at the annual all-hands meeting. This leader in the mobile Internet era seems to have lost its way in the AI era." The discussion of GPT will not begin until 2023, and the big model startups in the industry that are doing well are all founded between 2018 and 2021."

These words were like a thunderbolt that echoed in the tech world for a long time. While AI startups are catching up with you, the Internet giant, once known for its acumen, finally realized that it was lagging behind in this unprecedented technological change. Byte Jump's self-reflection, to some extent, also reflects the embarrassment of the entire Chinese technology industry in the field of AI:Once innovators, now they have to find the courage to catch up again.

This dramatic race has just kicked off. No one knows that a storm called Sora is brewing across the ocean as Chinese New Year approaches...

Chinese New Year Thunderbolt: The Shock of Sora

The New Year's Eve of 2024 has not yet arrived, the U.S. Silicon Valley has already sent a "special New Year's gift" to the Chinese AI circle.

On February 10th, the first day of the Lunar New Year, OpenAI released its literate video product Sora on social media, a model named after the Icelandic word for "story" that transforms text descriptions into high-quality videos up to 60 seconds long, supports multi-angle camera transitions, and even accurately reproduces complex physical laws of motion.

For a while, Chinese engineers who were supposed to be enjoying a reunion were forced to interrupt their vacations and nail themselves to their computers to study Sora's demo video.

This sudden "Spring Festival gift" has plunged the entire Chinese AI circle into contemplation. The emergence of this new track of video generation means another paradigm shift in technology. A CTO of an AI startup company who wished to remain anonymous exclaimed, "When we are still studying how to catch up with GPT-4, OpenAI is already opening up the next track."

March: Undercurrents

While companies are still digesting the impact of Sora, the capital markets are the first to react. At the end of FebruaryDark Side of the Moon Fires First Shot in Spring Funding Battle with Over $1 Billion in FundingIts luxurious lineup of investors - Ali, Sequoia China, Xiaohongshu, Meituan and Tonus Capital - demonstrates the capital market's firm confidence in Chinese AI companies.

March 6MiniMax also secured a $600 million funding round led by Ali, with a valuation that topped $2.5 billion in one fell swoop. The favor of capital seems to be preparing for the upcoming technology race

Almost at the same time, Zero One Everything released and open-sourced the Yi-9B model. This model, which excels in code and mathematical capabilities, quickly caused a stir in the open source community with its excellent English and Chinese language processing capabilities and low-cost deployment advantages.

On the product battlefield, AliCloud took the lead. on March 14, Tongyi Qianqian demonstrated impressive scene innovation capabilities, with its one-click parsing function to speed-read 10,000 pages of documents and the subsequent introduction of 6 hours of ultra-long audio/video transcription capabilities, demonstrating the results of deep plowing in the vertical field.

A few days later, the Dark Side of the Moon reported another success, as Kimi Intelligent Assistant increased the lossless context length to 2 million words, an achievement that far surpassed OpenAI and set a new benchmark for the industry.

In this March of uncertainty, a big announcement has galvanized the entire industry.

On March 19, NVIDIA announced at the GTC 2024 conferenceNew generation AI chip GB200This "super chip" is equipped with 208 billion transistors. This "super chip" with 208 billion transistors redefines the limits of AI computing with four times the training power of H100 and lower energy consumption.

The news gives both hope and pressure to the Chinese AI community, and the difficulty and cost of acquiring this top-tier arithmetic could polarize the industry.

March 23rd.The lowest-profile "AI Six Little Dragons" Step Leap Star finally unveils its mystery. The simultaneous unveiling of their Step series of generalized macromodel matrices - including the preview versions of Step-1, a giga-parameter language macromodel, Step-1V, a multimodal macromodel, and Step-2, a trillion-parameter MoE language macromodel.

At the end of the month, Ali once again showed the strength of technological innovation. the first open source MoE model Qwen1.5-MoE-A2.7B launched on March 29th, with 2.7 billion parameters can rival the 7 billion parameters of the traditional model, not only significantly reduces the cost of training, but also the industry to explore a new path to improve efficiency. This is another important breakthrough in Ali's model architecture innovation following the launch of the Qwen1.5 series before the Spring Festival.

From the shock of Sora in the Spring Festival to the release of NVIDIA GB200 in March to the hasty technical layout of various enterprises, this spring is destined to be unsettled. The temporary "loss of speech" of Chinese AI enterprises in the field of video generation and the urgent demand for top-level computing power all indicate that the industry pattern may undergo a drastic change. With the arrival of April, a multi-dimensional competition involving technology, capital and talent is quietly unfolding. ......

April: Hundred Boats

Entering April, the industry competition further heated up. on April 3, AliCloud launched theAI Programming Tools"Tongyi Spirit Code", which supports more than 200 programming languages, showed its ambition in the vertical field. 3 months later, this product wasWAIC (World Artificial Intelligence Conference) named as one of the town's treasures.

A week later, they launched the 32 billion parameter Tongyi Qianqi (Qwen1.5-32B), a model that balances performance, efficiency, and memory footprint to provide a more efficient and cost-effective solution for the industry.

April 17MiniMax releases abab 6.5 series of models, abab 6.5 containing trillions of parameters and the more efficient abab 6.5s. These two models support 200k tokens context length and can process nearly 30,000 words of text in 1 second, demonstrating amazing processing efficiency. More strategically, on April 29, MiniMax announced a comprehensive upgrade of its open platform API service, dropping the price by more than 50% and doubling the processing speed, taking an important step toward universal AI.

4月18日，Kimi智能助手迎来重大更新，包括模型能力提升、新增常用语功能、语音输入/播报以及搜索引用溯源，旨在提高用户体验和效率。新版Kimi在逻辑推理、数学编程、中英翻译等方面表现更优，同时支持个性化常用语设置和voice interaction，让信息获取更便捷。搜索结果新增引用溯源功能，确保回答的严谨性。

On April 30, an important evaluation result triggered the attention of the industry: Baichuan Intelligence's Baichuan 3 ranked the first in China with 73.32 points in the SuperCLUE Chinese large model evaluation, surpassing 32 large models such as GPT-4-Turbo. In particular, the outstanding performance in the two key dimensions of Knowledge Encyclopedia (82 points) and Logical Reasoning (68.60 points) has shown the industry the real strength of Chinese models.

May Fierce Battle: Giant Awakening and Price Storms

Competition on the AI track accelerated suddenly in May. on May 9, theAliCloud Releases Version 2.5 of Tongyi Thousand Questions and Open Sources 110 Billion Parameter Model Qwen1.5-110B. This version significantly improves comprehension and logical reasoning, and its Chinese language capability leads the industry. In several benchmark evaluations, Qwen1.5-110B successfully surpasses Meta's Llama-3-70B model, marking the first time that Tongyi's big model catches up with the GPT-4 level.

On May 13, OpenAI released GPT-4o ("Omni"), this all-in-one model is not only capable of analyzing and generating text, images and sound, it is twice as fast as GPT-4 Turbo and costs only half of the latter. For a while, the technology gap that Chinese companies had managed to close with great difficulty was once again widened.

Li Kaifu's leadership of zero one everything is also not willing to show weakness. on May 13, they released the 100 billion parameter AI model Yi-Large, and announced the open source closed source dual-track strategy, showing a clear and prudent business route.

On May 15, MiniMax launched its AI chatting app "Little Conch AI", which is based on a multimodal large model that can quickly process a large amount of text, understand emotions, and support multiple file formats and voice interactions.

In terms of industry governance, Smart Spectrum AI chose a unique entry point. on May 21, they signed the Frontier AI Safety Commitment with OpenAI, Google, Microsoft and 15 other top AI companies, joining the global AI governance dialog as an equal. On the same day, Baichuan Intelligence's Baichuan 4 set a new domestic record with 80.64 points in the SuperCLUE comprehensive benchmark evaluation, surpassing the GPT-4 Turbo, demonstrating the firm determination of Chinese companies to catch up in technology.

Followed by Baichuan Intelligence on May 22, launched a new generation of large models Baichuan 4 and the first AI assistant "Baixiaoying". This assistant in the general ability, math and code ability significantly improved, the ability of the domestic evaluation in the first place.

Two long-sleeping tech giants also woke up this month.

May 15Byte Jump releases a big model of the beanbag.Announced enough to stir the industry to use the price. The Beanbag general purpose model, pro-32k version, has a model inference input price of only $0.0008/thousand Tokens, while the pricing of the same size model in the market is generally $0.12/thousand Tokens, 150 times the price of the Beanbag model. pro-128k version has a model inference input price of $0.005/thousand Tokens, which is 95.81% lower than the industry's price. TP3T.

At the end of the month, Tencent also joined the fray. Tencent Yuanbao AI products are developed based on Tencent's hybrid big model with multimodal capabilities, aiming to provide instant answers, creative inspiration and fresh information, covering multiple scenarios such as knowledge learning, life encyclopedia, workplace office and fun creation.

However, not all is bright in this fierce competition. Unfortunately, after this heavy release of Baichuan Intelligence in April 2024, no similar follow up was seen in the following six months or so, and the year ended with news of the departure of co-founder and head of commercialization Hong Tao, casting a shadow over the quarter.

As the smoke of May fades away, the sunshine of June will shine into this track full of expectation and uncertainty. Who will come out on top in this unsmoked war remains an unresolved mystery.

Who fired the first shot in the price war?

On the fierce track of artificial intelligence, Byte Jump, a former sleeping giant, finally awoke, only to find that the world had turned upside down. If DeepSeek fired the first shot of the price war with that one price bullet, Byte was like a heavy gunner, instantly igniting an all-out price war.

At the beginning of May, DeepSeek, a subsidiary of the private equity giant Phantom Square, was like a dark horse, taking the lead in launching a price-cutting raid with the cost-effectiveness of the DeepSeek-V2 model. This model, which has approached the GPT-4 in terms of math, programming, and Chinese and English capabilities, uses a price of only 1/35 of the GPT-4o, instantly igniting the industry's sensitive nerves.

Byte jumping then joined the fray, in its usual aggressive style, the input price of the beanbag generic model Pro-32k was ruthlessly slashed to 0.8 yuan / million tokens. volcano engine president Tan to be even more boastful:"Large models are henceforth valued in centimeters."This move is tantamount to issuing a general mobilization order for a price war to the entire industry.

Ali Cloud, Baidu followed, have significantly reduced prices, and even directly launched a free model. KU Xunfei and Tencent are not willing to show weakness, Starfire Lite API and hybrid large model lite 256k announced free. In just a few days, the domesticAI macromodelThe market is already in an all-out price fight.

However, this seemingly fierce price war essentially exposes the deep anxiety of domestic AI enterprises.Price war is like drinking hemlock to quench the thirst, seemingly painful, but in fact, the crisis.Behind the companies' attempts to trade low prices for market share is confusion about business models and uncertainty about the future.

This move of byte jumping seems to be severe on the surface, but it reveals its helplessness on the AI track. As an Internet giant accustomed to winning through scale and traffic, they seem to have not yet found a real way to win in the field of AI. The price war is just one of the few weapons in their hands.

This price war without smoke is dragging all participants into a narrower and narrower channel. Technological innovation is drowned by the vortex of price, and commercial value is diluted by disorderly competition. Who can maintain rationality and long-term vision in this seemingly fierce but actually internal competition, who may ultimately stand in the high point of this emerging track.

When the smoke of the price war clears, what is left behind may only be a patch of chicken feathers and confusion about the future. On the AI track, which is destined to change the world, price should never be the ultimate winning formula.

Sixth, July dark surge: talent, capital and arithmetic of the game

The smoke of the price war has just cleared, and the AI industry has not ushered in calm, but has instead entered a more brutal playing field - the white-hot battlefield of talent, capital and arithmetic. If May is the stage for the awakening of the giants and the price storm, then June is a sign that a deeper competition is about to unfold.

In June, capital has an unusually keen sense of smell.On June 3, Smart Spectrum AI was the first to reap the benefits of international capital, securing $400 million in financing from Saudi Aramco's fund Prosperity7, the valuation exceeded 3 billion dollars, which is undoubtedly a strong endorsement of its technical strength and development prospects.

Immediately after that, on June 17, Hangzhou DeepSeek announced the open source of DeepSeek Coder V2, a model that is directly comparable to GPT-4-Turbo in terms of code and mathematical ability, with total parameters of 236 billion, ranking among the top in the world. The comprehensive open source of the model, code and thesis has set a benchmark of open sharing for the industry and accelerated the prosperity of China's AI ecosystem.DeepSeek's open source move not only demonstrates its technological self-confidence, but also attracts the attention of more developers and researchers, which lays the groundwork for the subsequent battle for talents.

If the flow of capital is surging in the dark, then the competition for talent is an open game. in July, the war for talent entered a white-hot. Byte jumping showed shocking offensive, with high specification treatment to dig in "the most understanding of Ali big model people" - the former Tongyi big model head Zhou Chang and his team, more netted zero one million things former vice president of algorithms Huang Wenhao, the original core members of the wall of the intelligence and so on. A number of top talents in the field of AI.

This series of heavy reinforcements, instantly rewrite the position of bytes in the AI talent map, but also heralded the Internet giant will launch a more fierce offensive in the field of AI. This kind of "poaching" behavior, although commonplace in business competition, but also reflects the extreme scarcity of talent in the field of AI and the thirst of enterprises for top talent.

Meanwhile, the situation of Zero One Everything is particularly difficult. Following the departure of technology co-founder Huang Wenhao, co-founder Li Xiangang also chose to return to the real estate trading platform Shell, and product head Cao Dapeng left immediately after.

The AI newcomer, which was founded just over a year ago, is facing the continued disintegration of its core team at a time when its valuation once reached RMB 20 billion. The near absence of heavyweight product releases in the second half of the year has cast a shadow over the star company's development and sparked industry concerns about the stability of talent at AI startups.

The capital frenzy has not been tempered by the movement of talent.At the end of July, Baichuan Intelligence completed RMB 5 billion in Series A financingThe top investors such as Ali, Xiaomi, Tencent, and CICC came out of the woodwork, proving once again that the capital market continues to be optimistic about the AI track.In August, the dark side of the moon received more than $300 million in financing again, and its valuation climbed straight up to $3.3 billion, consolidating its leading position in the field of AI.

In August, StepStar welcomed a heavyweight - Zhang Xiangyu, one of the four authors of ResNet. The joining of this 90-year-old AI bull subsequently attracted the followup of Yu Gang, Tencent's research director, and Duan Nan of Microsoft's Asia Research Institute. This series of talent intake not only brings technical boost to the company, but also sends a key signal to the industry: on the road to catch up, original technological innovation may be the real magic weapon to retain talent. Instead of high-paying poachers, it is better to build a more attractive technology platform and research atmosphere.

At the end of the year, millet in the field of AI has always been relatively "conservative" enterprises, also began to make efforts to ten million annual salary to dig DeepSeek V2, one of the core developers, the 95 genius girl Luo Fuli, to break the outside world in the field of AI stereotypes of insufficient investment.

At the same time, in the global capital market, another shocking story is being staged. on June 5, NVIDIA's share price soared, the market value for the first time exceeded 3 trillion U.S. dollars, surpassing Apple to become the second-highest global market value of the company.

By June 19th.More so, it surpassed Microsoft as the world's most valuable company with a market capitalization of $3.34 trillion.Over the past five years, its stock price has soared 3477.31%, far surpassing Microsoft and Apple, becoming the deserved arithmetic hegemony in the AI era, and highlighting the key role of arithmetic in the development of AI.

The flow of talent, the influx of capital, and the rise of the arithmetic hegemony together constitute the main theme of the AI industry in June. And the upcoming July will be an important stage for technology showcases and industry exchanges.

July Event: WAIC Stars Shine, DeepSeek Gains Notoriety

If June is the dark current, July is the stage to show strength and exchange ideas under the spotlight. The World Artificial Intelligence Conference (WAIC), held annually in Shanghai, is the highest-specification AI event in China, and the 2024 conference is even more unprecedented in scale, attracting global attention.

On the eve of the conference, June 13, StepStar's LeapQuest App was officially launched, integrating photo Q&A, intelligent search and other functions, aiming to improve work-study efficiency and simplify life. This app, built on Step Star's Step series of large models, optimizes networking search and document parsing capabilities, supports photo recognition and voice input, as well as document analysis in multiple formats, providing users with a convenient AI assistant.

On July 4, at the WAIC 2024 conference, theStep Star has released three new Step Series large models, including the Trillion Parameter Language Model, multimodal macromodels and image generation macromodels, and realized the leap from hundreds of billions to trillions of parameters, and made a breakthrough in the field of multimodality.

On the second day, Premier Li Qiang visited the 2024 World Artificial Intelligence Conference and toured the pavilion, paying a special visit to StepStar's booth. StepStar showed the Premier the latest progress of its Step series of generalized big models, including trillion-parameter language big models and multimodal understanding generation technology.

On July 5, Smart Spectrum AI also released at the World Conference on Artificial Intelligence theCodeGeeX fourth generation modelCodeGeeX4-ALL-9B model collects code completion, Q&A, interpreter and many other functions, and becomes the most powerful all-round code model with the strongest performance under 10 billion parameters.

In addition, on July 30, Kimi Intelligent Assistant launched a PPT creation tool to improve office efficiency.Since then, the PPT generation function has gradually become the standard for domestic AI tools.Kimi's move also reflects the penetration and popularization of AI technology in the office space.

The successful organization of WAIC not only demonstrated the latest progress of Chinese AI technology, but also promoted exchanges and cooperation in the field of AI at home and abroad. And shortly after WAIC, news from international authoritative evaluation organizations further affirmed the strength of Chinese AI technology.

On July 16 (US time), the results of the Big Model Arena update organized by LMSYS showed thatDeepSeek-V2-0628 Surpasses Several Top Models and Tops the List of Global Open Source Models. This achievement proves the competitiveness of China's open source big models on the global stage and earns international reputation for China's AI industry.

August: Gathering Momentum - Technology Innovation and Application Expansion

July's WAIC conference is still hot, the competition in the field of AI did not stop, but with the arrival of late fall, entered a more intense game stage. If it is said that the previous stage is the initial exploration and layout of the forces, then the late fall means that the real battle has officially opened the curtain. The flow of talent continues, technological breakthroughs are also endless, each enterprise is trying to find their own foothold.

In August, the eyes of capital remained focused on the AI track.Dark Side of the Moon raises over $300 million more in funding, valuation rises to $3.3 billion.This huge amount of financing undoubtedly injects strong confidence and financial support for the future development of this company, and signals that they will play a more important role in the next competition.

At the same time, technological breakthroughs have also begun to emerge. on August 6, Wisdom Spectrum AI achieved an important breakthrough in the field of video generation, open-sourcing CogVideoX video generation model. This lightweight model, which requires only 18GB of video memory to achieve 6 seconds of video generation, greatly reduces the threshold for developers to use, allowing more people to participate in AI video creation.

More surprisingly, just less than a month later, on August 28th, the CogVideoX-5B model with larger parameters and stronger performance was also announced as open source, and the video memory requirement was lowered to a minimum of 11.4GB. The successive breakthroughs of Wisdom Spectrum AI in the field of video generation not only show its strong technical strength in this field, but also accelerate theAI Video GenerationPopularization of technology.

Not only that, byte jumping also launched a one-stop AI creation platform called "namely dream AI" on August 6, directly targeting the fast hand of Keling and Sora, to further expand its layout in the field of AI creation, in an attempt to occupy a more favorable position in this emerging market.

On the technical level, DeepSeek also significantly reduced API service latency and cost through innovative hard disk caching technology on August 2, significantly improving the user experience and laying the foundation for subsequent larger-scale applications.

All in all, August was a month of technological innovation and application expansion going hand in hand to build momentum for the competition to come.

September: A Hundred Flowers - Multi-disciplinary Breakthroughs and Evolution of the Landscape

Entering September, various enterprises have launched more active exploration at the level of technology and application, and the competition has become more white-hot, showing a blossoming situation.

On September 6, Wisdom Spectrum announced that its AI product "QingYinVideo calling is now fully availableThis new feature breaks through the limitations of traditional typing and voice interaction, enabling AI to "see" the world and understand user expressions and emotions, thus providing a more natural and smooth interaction experience. This new feature breaks through the traditional limitations of typing and voice interaction, enabling AI to "see" the world and understand users' expressions and emotions, thus providing a more natural and smooth interaction experience, which undoubtedly raises the user experience to a new level. This marks an important progress in multimodal interaction of the big model of Smart Spectrum, and it also successfully catches up with the level of GPT-4o released by OpenAI in May, which demonstrates the speed of Chinese AI enterprises in catching up with technology.

Also on September 6, DeepSeek released its V2.5 model, which not only integrates general-purpose conversation and code processing capabilities, but also significantly optimizes human preference alignment, writing and command following, and continues to maintain practical features such as Function Calling, FIM Completion, and Json Output to improve the model's comprehensive performance.

DeepSeek V2.5 has lived up to its reputation and won the top spot in the subsequent global big model arena, ranking the first in China, even surpassing the strongest closed-source model in China, and leading the domestic models in 8 individual capabilities, proving its strong technical strength to the world once again, and also winning international reputation for Chinese open-source models.

On September 10, Kimi API started to support the connected search function, becoming the first Chinese AI company to launch a function similar to OpenAI Search, providing users with a more convenient and intelligent conversation experience, and setting a new benchmark for other companies to promote the development of AI applications.

What's more, on the same day, Apple officially launched "Apple Intelligence" at its fall event, an event of epoch-making significance.It marks the official entry of AI at the level of cell phone operating systems, opening a new era of AI phones.Apple Intelligence is deeply integrated into iOS, bringing users unprecedented intelligent experiences such as smart notification summaries, email auto-replies, and smart editing of photos.

This move quickly triggered a shock in the entire cell phone industry, and in the following months, it triggered a collective follow-up by Chinese cell phone manufacturers, who launched AI OSes that benchmarked Apple Intelligence in an attempt to seize the lead on the new track.The release of Apple Intelligence is undoubtedly one of the most important events in the cell phone industry in 2024. It not only changes the way users interact with their phones, but also opens up new application scenarios for the development of AI technology.

On September 12, OpenAI launched o1-preview and the faster and cheaper o1-mini, once again pointing out a new direction for the industry, both of which emphasize more on investing more "thinking time" before answering to improve the ability to solve complex problems, providing a new way of thinking for the development of large models.

More importantly.The launch of OpenAI o1 marks the formal entry of AI development into the "reasoner" stage.Whereas previous AIs were more "executors", able to complete tasks according to instructions, o1 is beginning to show some reasoning ability, able to better understand the problem, analyze the information, and give more reasonable answers.

Once again, Chinese companies see new targets and are actively exploring technological breakthroughs in the direction of "reasoning" in an effort to take the lead in the next wave of AI technology.

September is also a key month for the video generation track.MiniMax's conch video generation model abab-video-1 earns attention at home and abroadIt has not only gained popularity among domestic netizens, but also reaped extremely high praise among foreign users, demonstrating the potential of Chinese AI in the field of video generation.

However, it is regrettable that Zhang Qianchuan, MiniMax's product leader and the helmsman of "Hoshino" and "Talkie", also stepped down from the company's affairs this month due to personal reasons, and changed to the position of product consultant, which undoubtedly adds a touch of uncertainty to MiniMax's future development. This undoubtedly adds a trace of uncertainty to the future development of MiniMax, and also triggers the industry to think about the stability of the talent of AI startups.

On September 19, AliCloud's Tongyi Wanxiang was officially unveiled at the Yunqi Conference and showed its unique advantages in a variety of styles such as Chinese style, 3D animation and CG thick paint, attracting a lot of attention and providing more possibilities for AI art creation.

At the same conference, AliCloud even announcedQwen 2.5-72B model open-sourced globally and announced to outperform Llama 405B, which supports 128K tokens and generates 8K tokens of content, fully demonstrates the great breakthrough of AI in programming and multimodal capabilities, and further promotes the development of the open source ecosystem.

On September 20, Tencent Yuanqi AI Intelligent Body was officially released, bringing new possibilities for public number creation, and marking a further deepening of the application of AI in the field of content creation, heralding a change in the way content is produced.

September 24Byte Jump has also released the Beanbag Video Generation Mega Model thatAnd claimed that it breaks through the multi-subject interaction difficulties, supports multi-style multi-scale consistency multi-camera generation, applicable to e-commerce marketing, animation education and other fields, will undoubtedly further intensify the competition of the video generation track, and promote the technological progress and application innovation in this field.

On September 25, Baidu AI's Wenxin Express Code won the first place in both Sullivan and SuperCLUE's two authoritative evaluation reports, and took the lead among domestic AI code products with a total score of 87.55.

September can be described as a month of blossoming, with companies making impressive progress in different directions.

From August to September, China's AI industry has been booming in terms of technological innovation, application expansion and talent mobility. All companies are actively exploring their strengths and breakthroughs, and together they are driving the progress of China's AI industry. What new stories will happen in the next few months?

Video generation: from catching up to surpassing - China's AI breakout road

Among the many branches of AI technology, video generation is undoubtedly one of the most popular focuses in recent years. On this track full of challenges and opportunities, Chinese AI companies have experienced a journey from catching up to surpassing.

The shocking release of OpenAI's Sora during the Chinese New Year brought a huge impact to the global AI community, and once made Chinese AI companies feel the pressure. However, this pressure has instead inspired Chinese companies to innovate and catch up.

Just a few months later, Chinese companies proved themselves in action.6On June 6, Racer took the lead in launching its self-developed video generation model "Keling".. The product demonstrated stunning power right out of the gate: 1080p Ultra HD resolution, the ability to generate up to 2 minutes of video, and the freedom to adjust the aspect ratio - all key metrics that were dramatically ahead of the industry at the time, and even surpassed Sora, which had yet to be officially released at the time.

The development trajectory of "Keling" can be described as steady: the launch of graphic video in June, the opening of the web terminal in July, and the launch of the "AI director co-creation program" and version 1.6 in December. Its AI-generated movie and TV dramas and other content have been screened on major social platforms, firmly occupying the position of leader in the field of video generation.

In September, MiniMax's conch video generation model abab-video-1 rose to prominence, not only reaping positive reviews at home, but also gaining high recognition among overseas users. Meanwhile, Vidu, Pixverse and other startups have also shown excellent technical strength. Tencent's open source hybrid video model even surpasses Sora in terms of effect.

When OpenAI finally released Sora after nearly 10 months of waiting, it brought unexpected disappointment to the market. Due to various reasons, the actual effect of Sora is far from the initial demo video, not only lagging behind Google's Veo2, but also left behind by many Chinese products. This marks the first time that a Chinese company has truly surpassed OpenAI on the video generation track.

The success in this track has given great confidence to Chinese AI companies. in July, Wisepac AI released "清影", which generated millions of videos within 6 days of its launch. In order to maintain the competitive advantage, Zhipu quickly open-sourced the CogVideoX model in August, and in November, "qingying" was upgraded to support 4K, 60 frames of ultra-high-definition video generation and added the CogSound sound model. in September, AliCloud's Tongyi Wanxiang chose to seek breakthroughs in the vertical areas of national wind and 3D animation.

The success of the video generation track not only proves that Chinese AI companies have the strength to outperform international giants in niche areas, but more importantly, that theIt breaks the curse of "catching up forever", injects new confidence into China's AI industry, and signals that China's AI will move towards a new stage of independent innovation and leadership..

In the aftermath of the price war, Chinese AI companies are beginning to step into a deeper contest: the track of technological innovation and global competition.In November, this battle without smoke quietly heated up, and every tiny breakthrough may redefine the industry ecology.

November: Accelerated period of technological innovation

As the year draws to a close, Chinese AI firms embark on a final sprint to 2024.

On November 19, Step-2 of Step Star ranked fifth in the world in the international authoritative list LiveBench, second only to OpenAI's o1-mini, an achievement that signifies that the strength of Chinese AI enterprises in the international arena is gradually improving. During the same period, its Step-1V was ranked No. 1 in China in Chatbot Arena's latest list, alongside Gemini-1.5-Flash, demonstrating impressive technical strength.

Tencent has taken the lead in open-sourcing models and multimodal applications, and on November 5, Tencent officially open-sourced its hybrid large language model and 3D model. Its newest MoE model, "Mixed Elements Large", has a parameter scale of 389B, which is a leader in multidisciplinary evaluation.

"Hunyuan3D-1.0″ supports text-image generation in 3D, providing a powerful tool for developers and researchers.On November 14, Tencent Yuanbao 2.0 was fully upgraded, with a new exclusive section for AI applications, and a hybrid model architecture that supports multimodal understanding and generation, further expanding the application boundaries.

However, the road of technological innovation is not smooth sailing. November 19, Tencent hybrid big model technology leader Liu Wei chose to leave, this personnel change triggered the industry's attention to the flow of talent.

Meanwhile, Baidu showcased new technological breakthroughs at its World Congress. Robin Li announced the launch of iRAG, a retrieval-enhanced graphic technology, and "Seconda", a no-code tool dedicated to solving the problem ofAI image generationThe problem of illusion in the "second da" allows non-programmers to easily realize the creativity, marking the AI application is moving towards the masses.

In terms of mathematical and reasoning capabilities, Kimi intelligent assistant released a new generation of algebraic reasoning model k0-math on November 17, whose mathematical problem solving capabilities are benchmarked against the OpenAI o1 series. The Kimi Exploration Edition launched at the same time enhances search intent, source analysis and chain thinking capabilities to provide users with smarter problem solutions.

November 20th.DeepSeek's new inference modelDeepSeek-R1-Lite Preview Released, which users can experience through the official website. The model performs well in the fields of mathematics and programming, the reasoning process includes reflection and verification, and the chain of thoughts can be up to tens of thousands of words in length, demonstrating a reasoning performance that exceeds that of models such as GPT-4o. Currently, it only supports web use, and will be open-sourced and provide API services in the future.

Throughout November, the common goal of Chinese AI companies seems to be very clear: to catch up with OpenAI's September release of o1 before the Chinese New Year. Baidu wenxin yiyin user scale has reached 430 million.AliCloud's QVQ-72B-Preview matches OpenAI o1 and Claude3.5 Sonnet for the first time in visual understanding and reasoning capabilitiesThe progress has confirmed the determination of domestic companies to catch up.

From technology reviews to model open-sourcing, from multimodal applications to reasoning capabilities, China's AI scene in November showed unprecedented activity and competition. Companies are closing the gap with international giants at an unprecedented rate, showing exciting potential for innovation.

This month's signs indicate that Chinese AI companies are no longer satisfied with imitation, but have begun to take the initiative to make their voices heard on the global stage.In December, the competition will enter a more intense stage.

December: a full breakthrough in innovation

If November is the prelude to the accelerated catching up of Chinese AI enterprises, then December is the key chapter of a full-scale breakthrough. This month, Chinese AI companies showed unprecedented aggressiveness in technological innovation, model development and commercial layout.

Step Star became the focus of this month. on December 13, the company launched Step-1o, the first end-to-end voice big model with hundreds of billions of parameters in China, which not only supports mixed input and output of voice and text, but also has a high IQ and EQ, and is able to understand the emotional information and provide professional advice and emotional companionship.

The launch of Step-1o signifies that the latecomer has fully benchmarked OpenAI's GPT-4o, released in May, and achieved a major breakthrough in the field of voice interaction. Immediately afterward, the company completed hundreds of millions of dollars in Series B financing, with a lineup of investors including Tencent Investment, Wogen Capital and Qiming Venture Capital, underscoring the capital market's confidence in the potential of its technology.

Kimi's Intelligent Assistant released Visual Thinking Model k1, a breakthrough model based on reinforcement learning technology, on December 16. k1 supports end-to-end image understanding and thought chaining technology, covering basic sciences such as math, physics, and chemistry. In many benchmark tests, the k1 model surpasses global benchmark models, giving Kimi wings to take off in the field of visual thinking.

DeepSeek intensively launched a series of heavyweight models in December. on December 10, the final version of V2.5 fine-tuned model was released, which improved the ability of math, code, writing and other dimensions through Post-Training. on December 13, DeepSeek-VL2 was formally unveiled, which introduced the dynamic cutting strategy and MoE architecture, and realized a significant improvement in the visual ability. On December 26, DeepSeek-V3 was launched, with 671B parameters, and excelled in multiple domain evaluations, especially in math and Chinese language capabilities, with the generation speed increased to 3 times.

Byte jumping in this month continued to force the AI ecological. December 4, bean bag AI assistant new picture understanding function, allowing users to upload pictures and get content analysis. December 11, the company elevated the priority of namely dream product, committed to creating "AI era of jittery voice." December 19, the company's AI assistant, the company's AI assistant, the company's AI assistant, the company's AI assistant, the company's AI assistant, the company's AI assistant.It's even rumored that talks are underway with Apple about plans to integrate its AI models into Chinese-market iPhonesThis news, if true, will be a major breakthrough in cross-border cooperation.

On the last day of the year 2024, Wiseplan Technology has delivered an amazing answer sheet: GLM-Zero preview not only catches up with OpenAI's o1, but also makes innovative attempts in reasoning methods. This model based on Extended Reinforcement Learning (ERL) technology is comparable to o1-preview in several evaluations, marking the transformation of Chinese AI enterprises from "followers" to "innovators".

Throughout December, Chinese AI companies seem to have found a delicate balance: while catching up with OpenAI, they have started to establish their own technical characteristics and innovation paths. From voice interaction to visual thinking, from multimodal modeling to inference technology, these breakthroughs are not just iterations of technology, but also the exploration of a brand new technology paradigm.

When the curtain falls on the last day of 2024, Chinese AI companies are already standing at a brand new starting point.In 2025, this global AI competition without smoke will be even more confusing.

2024: Dilemmas and Breakthroughs for Catch-Ups

Looking back to 2024, Chinese AI companies have traveled a journey full of ups and downs. The journey of catching up, which began at the beginning of the year, has experienced the impact of Sora, the challenge of GPT-4o, and the new goals brought by the o1 series. In this endless catching up, Chinese companies have shown amazing execution and rapid iteration capabilities, and every OpenAI innovation has been exchanged for a rapid response from Chinese companies.

However, this "you make a move and I'll follow it" model also reveals the lack of original breakthroughs. In terms of basic modeling and product innovation, Chinese companies are playing the role of "catching up". If anyone breaks this cycle in 2024, it may be DeepSeek's several original attempts. This low-profile company has not only cultivated AI wizards such as Luo Fuli, but also continued to make efforts in basic research, showing a different innovation path.

Looking ahead to 2025, Chinese AI companies are facing an even greater challenge: how to maintain the ability to catch up quickly while fostering a truly innovative soil. Byte Jump and Xiaomi's large layout in the talent market, Step Star's introduction of top scientific research talent, and Smart Spectrum AI's innovative attempts at reasoning technology all signal that the industry is experiencing transition pains. From "catching up" to "surpassing", the road may still be long, but the direction has become clearer.

Originally republished from: AI Fan, Original title:TheWhat has happened to AI in China in 2024? | The Road to Catch Up and SurpassThe