news

한어Русский языкEnglishFrançaisIndonesianSanskrit日本語DeutschPortuguêsΕλληνικάespañolItalianoSuomalainenLatina

written by | zhang yu

editor | yang yong

recently, ai big model "unicorn" zhipu ai completed a new round of financing led by zhongguancun science park company with a pre-investment valuation of 20 billion yuan, with a financing amount of tens of billions of yuan. it is reported that this round of financing is the third round of financing completed by zhipu ai since 2024.

in june 2024, it was reported that prosperity7, a fund managed by the venture capital arm of middle eastern oil giant saudi aramco, invested us$400 million in zhipu ai. this round of financing directly helped zhipu ai's valuation enter the "20 billion yuan club", and the 20 billion yuan valuation is also generally considered to be the baseline for entering the first echelon of the industry.

as of now, the list of shareholders of zhipu ai includes well-known investment institutions such as hillhouse capital, qiming venture partners, and sequoia capital, as well as internet giants such as meituan strategic investment department, ant group, alibaba, and tencent investment. in addition, there are state-owned forces such as the social security fund zhongguancun fund, beijing artificial intelligence industry fund, and zhongguancun science city.

zhipu ai was founded in june 2019 and was transformed from the technological achievements of the knowledge engineering laboratory (keg) of tsinghua university. it is currently the only fully domestically-funded and fully self-developed large-model enterprise in china. as early as 2020, it began the research and development of the glm pre-training architecture and trained the 10 billion parameter model glm-10b; in 2022, it cooperated to develop the 130 billion-level ultra-large-scale pre-trained general model glm-130b; in 2023, zhipu ai launched the glm series of 100 billion open source base dialogue models, and launched glm-4 in january 2024. since then, zhipu ai released the glm-4-9b open source model in june, and launched the video generation tool qingying in july.

"catching up with openai" and "benchmarking open ai is the goal of zhipu ai" are slogans that zhipu ai ceo zhang peng has repeatedly mentioned when sharing with the outside world. however, the current competition for big models is no longer a battle of having or not from 0 to 1, but a battle of implementation. with the support of many capital forces, can zhipu ai, known as the "chinese openai", successfully break through in the big model competition?

1. openai is still the leader

since its establishment, zhipu ai has always regarded openai as its target. so far, zhipu ai has created a complete model product that is comparable to openai, including the ai ​​efficiency assistant zhipu qingyan, the high-efficiency code model codegeex, the multimodal understanding model cogvlm, and the text graph model cogview, etc.

however, although zhipu ai claims to be china's first open source large model, it is not easy to compete with openai.

for example, zhipu ai released a new generation of language model glm-4 at the first technology open day (zhipu devday) held in january 2024. although the overall performance of glm-4 has been greatly improved by 60% compared with the previous generation, it is claimed to be "on par with gpt-4", but in fact it only reached about 90% of the level of gpt-4.

zhang peng also admitted that compared with foreign large models, the development of domestic large models started a little late. coupled with the limitations of high-performance computing power and the gap in data quality, domestic large models have a certain gap with the world's advanced level in terms of scale and core capabilities. this gap is about one year.

first, from the perspective of technical route, openai pays more attention to versatility, portability and scalability. its gpt series models can be applied in multiple scenarios and are highly customizable. in contrast, zhipu ai's technical route is "big model + small model", which adapts to the needs of different scenarios and tasks through pre-training and fine-tuning of large models. this technical route can improve the generalization ability and application scope of the model, but there are also problems such as high model complexity, large amount of calculation, and long training time.

secondly, openai's gpt series models are large in scale and can process large amounts of natural language data, resulting in better model performance. in contrast, zhipu ai's models may be smaller in scale and have limited data processing capabilities, which may affect its model performance and generalization capabilities.

in addition, in terms of data resources, openai has a large amount of natural language data resources that can be used to train and optimize its models, while zhipu ai's data resources may be relatively small, resulting in limited effects and performance of its model training.

the gap between the two sides is most directly reflected in the number of users. in november 2022, openai's chatgpt had more than one million users in just five days after its launch. in january 2023, its monthly active users exceeded 100 million, making it the fastest growing consumer application in history. in contrast, as of november 2023, the daily active user number of zhipu qingyan, a subsidiary of zhipu ai, ranged from only 100,000 to 400,000.

in fact, the gap between zhipu ai and openai is getting bigger and bigger. on september 13, openai released the o1 series of models, including the o1 preview version and o1-mini. in a series of benchmark tests, o1 has once again made a huge improvement over gpt-4o, and even "comparable to human experts" in benchmark tests of physics, biology, and chemistry problems.

for example, in the international mathematical olympiad (imo), gpt-4o scored 13.4%, while o1 scored 83.3%. in the codeforces programming competition, o1 scored an excellent score of 89%, while gpt-4o's accuracy was only 11%. in addition, in the gpqa-diamond test, the accuracy of human experts was 69.7, while o1 was as high as 78%.

it can be seen that zhipu ai is still far away from openai. although the achievements made by zhipu ai are already very rare, facing openai's newly launched o1 series models, zhipu ai undoubtedly needs to work harder.

2. price war intensifies

since may 2024, the price war in the large model field has been going on for more than four months, causing more and more large model companies to be caught in the vortex of price wars.

the price war started with deepseek, an ai company under the private equity giant huanfang quantitative. on may 6, deepseek announced the open source of the second-generation moe large model deepseek-v2, priced at nearly 1% of gpt-4-turbo, and only 1 yuan for one million tokens.

zhipu ai followed closely. on may 11, zhipu ai announced that the calling price of the personal version of glm-3turbo was reduced from 5 yuan/million tokens to 1 yuan/million tokens. at the zhipu ai open day event held on june 5, zhipu ai once again announced a price reduction for the entire model matrix. among them, the price of glm-4-air and glm-3-turbo was reduced to 0.6 yuan/million tokens, the embedding-2 model was as low as 0.3 yuan/million tokens, and the price of the glm-4-flash model was reduced to 0.06 yuan/million tokens.

bytedance has also joined the price war, announcing that the doubao main model (doubao universal model pro) will be priced at 0.0008 yuan/thousand tokens in the enterprise market, under the banner of 99.3% lower than the industry average price. the price of the same specification model on the market is generally 0.12 yuan/thousand tokens, which is 150 times the price of the doubao model.

since then, alibaba, tencent, baidu, and iflytek have all announced price cuts for large models. for example, alibaba cloud has reduced the input price of qwen-long to 0.0005 yuan/thousand tokens, and the output price has dropped by 90% to 0.002 yuan/thousand tokens; baidu smart cloud has announced that the two main models of wenxin large model, enire speed ​​and enire lite, are free to use.

openai is also the main force in the price war. the price of its gpt-4o has been halved again compared to gpt-4-turbo. this is the fourth price cut by openai since the beginning of 2023. according to openai's expectations, its large models will continue to decline by 50%-75% per year.

it is worth mentioning that although the continued decline in large model pricing is expected to bring about faster commercialization, price wars often mean that large model companies need to make concessions on price. for zhipu ai, its own profitability is already limited. if the price war continues, it may lead to further declines in profits, and it will become more difficult to achieve profitability.

in comparison, the price war may have less impact on openai. after all, as early as december 2023, openai ceo sam altman revealed that openai's current monthly revenue has reached hundreds of millions of dollars, and its annualized revenue is likely to exceed 1.5 billion. third-party institutions also predict that openai's revenue in 2024 is likely to more than double that of 2023, and optimistic estimates will reach 5 billion us dollars.

it is foreseeable that with the price war and the technological gap, zhipu ai may not have an easy time in 2024. zhang peng also admitted that the challenges faced by zhipu ai in 2024 are very arduous: on the one hand, openai will achieve new breakthroughs in super cognition and super alignment technologies in 2024, which requires zhipu ai to continuously iterate its technology and keep up with the world's leading pace; on the other hand, large models will usher in a wave of commercialization in 2024, and zhipu ai's commercial competition pressure will also increase.

3. accelerate ecological investment

product layout and investment layout are the two main lines for zhipu ai to achieve commercialization.

zhang peng once publicly explained zhipu ai's investment strategy: "we hope to build a large model ecosystem, in which we work together with partners to make the ecosystem bigger and bigger. this is our longer-term commercialization goal." when talking about the business vision for 2024, zhang peng said: "it is our important task to make the big model truly grounded and down-to-earth."

in 2024, zhipu ai will launch an open source big model open source fund, which includes three "1000s": zhipu ai will provide 1,000 computing cards to the big model open source community to help open source development; provide 10 million yuan in cash to support open source projects related to big models; and provide 100 billion free api tokens to outstanding open source developers. zhang peng said that the purpose of the big model open source fund is to promote the great progress of big model research and development and promote the great prosperity of the entire open source ecosystem of big models.

facing the global big model entrepreneurs, zhipu ai will upgrade the "z plan" and jointly launch a big model entrepreneurship fund with ecological partners with a total amount of 1 billion yuan to support original innovation of big models, covering big model algorithms, underlying operators, chip optimization, industry big models and super applications.

in august 2024, at the zhipu ai "z plan" corporate roadshow event, zhang peng officially announced that zhipu ai and its ecological partners would launch the agi ecological fund: z fund to support more early-stage projects with potential in the large model track. on september 3, humanoid robot manufacturer dynamics technology completed an angel round of financing of tens of millions of yuan, led by z fund, which was also the first external investment of z fund. so far, zhipu ai has invested in 11 companies, including ai model layer company "lingxin intelligence", intelligent legal service product provider "power law intelligence", software and information technology service provider "shudao zhisuan", generative ai application provider "shengshu technology", etc.

objectively speaking, in the case of insufficient supporting industrial chain, investing in the layout of the entire industrial chain is a way to break the deadlock, but the more critical way to break the deadlock lies in how to create products that expand users' imagination and how to turn these products into productivity. this will be the next must-answer question for zhipu ai.