Coding Questions Everything: The Two-Year Survival Struggle in the Independent Large Model Race (Part 1)
If you were to summarize what has happened in the independent large model track in the past two years in one sentence, it could be like this:
The label "Six AI Dragons" was born in April 2023 and died in the second half of 2025.
One fact is that the six companies have become four.
Zero One Everything abandoned the pre - training of ultra - large foundation models with over a trillion parameters in 2025, and Baichuan Intelligence made its turn earlier. In the all - staff letter on the second anniversary of its establishment on April 10, 2025, Wang Xiaochuan admitted: "In the past two years, Baichuan had an over - extended front. Next, we will shrink the front and bet on medical AI."
The remaining four companies must re - justify their existence.
But the challenges came much faster.
On January 20, 2025, DeepSeek - R1 was released. Seven days later, NVIDIA's stock price dropped by 16.97% in a single day, and its market value evaporated by about $590 billion, setting a record for the single - day market value loss of a US listed company. But the real gap torn open by DeepSeek that week was not NVIDIA's stock price, but the core narrative of the Chinese large model industry. When a company not on the list of the Six Dragons made an open - source model comparable to the top - tier closed - source models with 1/100 of the training cost, all the money - burning, advertising investment, and soaring valuations around the "Chinese version of OpenAI" in the past two years need to be re - justified.
This is not easy. At the beginning of 2026, the remaining four companies completed the most important identity transformation in the independent large model track in the past two years on the Hong Kong stock market. Zhipu rang the bell on January 8, and its Hong Kong public offering was subscribed 1159.46 times. MiniMax rang the bell on January 9, with a subscription of 1837.17 times, setting a new record. The Dark Side of the Moon was reported by LatePost to have completed a $2 billion Series D financing in May, and its post - investment valuation exceeded $20 billion. Step Star was reported by National Business Daily to be about to complete a nearly $2.5 billion financing, with a target cornerstone pricing in the range of tens of billions of dollars.
It seems that everyone has a bright future, but in fact, the market value gap among the four companies is much larger than what the unified label "Four AI Dragons" implies. Although exchange rates may cause some caliber issues, basically, Zhipu's market value is twice that of MiniMax. Although the valuation of The Dark Side of the Moon has soared, it is still only a little over 30% of Zhipu's. Step Star has the seemingly most luxurious technical lineup, but its target valuation is the lowest.
Behind this gap is the structural change when the real productivity battlefield in the Chinese large model industry shifted from "super applications" to "Coding". This shift shaped the fate of each company at different times and to different extents.
And the source of this shift is a report released in December 2025.
——Introduction
What 100 Trillion Tokens Tell Us
On December 4, 2025, OpenRouter and a16z jointly released a report titled "The State of AI: An Empirical 100 Trillion Token Study with OpenRouter". This report is based on the real usage data of over 100 trillion tokens on the OpenRouter platform in the past 13 months.
There is a set of data in the report that completely shattered the "super - application illusion" that the Chinese large model industry debated about throughout 2025. The proportion of programming in the token usage of paid models soared from 11% at the beginning of 2025 to over 50% in November 2025.
This means that all the battlefields that everyone was involved in the Chinese large model industry in the past two years, such as long - text thinking, daily active users of AI assistants, customer acquisition costs through advertising investment, and soaring valuations, are not the real token economics engine of this revolution. The real engine is Coding.
From this perspective, China's leading large model companies still have a gap compared with the world's top - tier models. Anthropic's Claude once occupied more than 60% of the share in the API call volume of programming. This figure itself is a heavy blow. It tells all Chinese independent large model companies that the global first - tier in the Coding main battlefield is still actually occupied by US closed - source giants.
But Chinese companies are not absent either.
Open the mainstream Coding evaluation lists from 2025 - 2026, such as SWE - bench Verified, SWE - Bench Pro, Terminal Bench 2.0, Multi - SWE - Bench, SWE - rebench. Chinese names will appear repeatedly. Alibaba's Qwen3 - Coder series is close to 70% in SWE - Bench Verified and has long been at the forefront of open - source in SWE - rebench. DeepSeekV3, R1, V3.2 are continuously on the list. Among the Four Dragons, the companies that bet on Coding the earliest and deepest, such as Zhipu GLM - 5, Kimi K2.5, MiniMax M2.5, are also on the list.
Although Alibaba's Qwen3 - Coder does not belong to the category of "independent large models" discussed in this article, its existence precisely proves one thing: Coding is the real main battlefield. Even an Internet giant with an annual revenue of over a trillion dollars is going all out on this.
Whether independent large model companies can get a position on this main battlefield directly determines their market value rankings. The market value differences among Zhipu, The Dark Side of the Moon, MiniMax, and Step Star in May 2026 almost perfectly correspond to the time sequence, investment depth, and decision - making styles of their respective bets on Coding.
To understand this, we have to start with four people.
Four Generations of People
On the afternoon of January 8, 2026, at the ringing - the - bell ceremony at the Hong Kong Stock Exchange, Zhang Peng, the CEO of Zhipu AI, stood on the stage.
But Zhang Peng is not the person with the strongest "founder feeling" in this company. The real soul figure of Zhipu is Tang Jie, a professor in the Department of Computer Science at Tsinghua University and a top - tier scholar in the field of knowledge graphs and graph neural networks in China. He was born in 1977 and is exactly 49 years old in 2026. He did not appear on the ringing - the - bell stage, but he is the core of all decision - makings for this company from the Tsinghua KEG Laboratory to the Hong Kong Stock Exchange.
This 49 - year - old Tsinghua professor led a company with an academic background to go public at the darkest moment in the Chinese large model industry, and then forced the entire industry to accept that "AI should not be free" with the first - place ranking in the open - source Coding evaluation.
24 hours later, MiniMax rang the bell at the Hong Kong Stock Exchange. On the stage was Yan Junjie, born in 1989 and 32 years old when he started his business. When he left his position as the vice - president of SenseTime, everyone thought he would continue on the path of the "Chinese version of SenseTime", targeting the B - end, government and enterprise sectors, and computer vision. But he made the opposite choice: native to the C - end, born global, and parallel in all modalities.
On February 12, 2026, the Chinese large model industry witnessed the most dramatic day in its history. Two direct competitors released flagship models of the same type on the same day.
This was not a coincidence. Zhipu and MiniMax had respectively determined around February 12 as the release window for their new models as early as Q4 2025. Both companies knew that the other would release at this time. But neither was willing to be the first or the last to release. Releasing first would be compared with the other, and releasing last would be considered as passive following.
The final result was that they released on the same day, almost at the same hour.
Zhipu chose the early morning. GLM - 5 was officially released at the early morning of February 12, 2026. MiniMax chose the morning. M2.5 was released on the same morning. This time difference between "early morning" and "morning" is itself a competitive strategy. Zhipu released in the early morning to seize the early - morning attention of overseas media, and MiniMax released in the morning to seize the top headlines of domestic media on the same day.
Within 24 hours after the release, the official accounts, cooperative media, and KOLs of the two companies almost flooded the screens at the same time. Zhipu focused on "first in open - source Coding and price increase", while MiniMax focused on "one dollar per hour, cost revolution". The grand occasion of the two companies flooding the screens at the same time only occurred when DeepSeek - R1 was released in early 2025.
Yan Junjie felt heavy - hearted. He is 37 years old, from a technical executive position in a large company. He led a company with a "reverse bet" to get 1837 times of subscriptions at the Hong Kong Stock Exchange, but its market value is only half of Zhipu's.
And Yang Zhilin did not ring the bell at the Hong Kong Stock Exchange. He is still waiting. He was born in 1992, graduated from the Department of Computer Science at Tsinghua University with a bachelor's degree and obtained a doctorate from Carnegie Mellon University. During his time at Tsinghua, he passed all programming courses with full marks and graduated first in his class. At CMU, he obtained a doctorate in only 4 years, which usually takes 6 years. This background is almost unique among the founders of Chinese independent AI companies.
In 2024, he was called "the youngest star entrepreneur in Chinese AI" by the media. In the middle of 2025, after DeepSeek emerged, he disappeared from the public opinion field for 172 days. On March 19, 2026, he gave a 40 - minute technical speech in fluent English on the main stage of the NVIDIA GTC Conference. He is the only founder of a Chinese independent large model company invited to give an on - site speech at this GTC. Elon Musk commented on this speech on social media as "impressive".
This 34 - year - old genius led The Dark Side of the Moon to complete 3 rounds of financing in 3 months, and its valuation soared from $4.3 billion to $20 billion. But Yang Zhilin's biggest challenge is that he himself is almost the entire supporting wall of the company's technical narrative, and he does not have a team with sufficient "depth on the bench".
Step Star's situation is the most special. Jiang Daxin, the CEO, is a former global vice - president of Microsoft, born around 1976 and elected as an IEEE Fellow in 2025. He is one of the most academic and senior technical decision - makers among Chinese independent large model companies. But in January 2026, Step Star announced that Yin Qi, the former co - founder of Megvii Technology and the chairman of Qianli Technology, would serve as the company's chairman.
Yin Qi was born in Wuhu, Anhui in 1988 and graduated from the Yao Class at Tsinghua University, studying under Turing Award winner Yao Qizhi. He founded Megvii Technology at the age of 23. He is one of the operators of the Four Dragons in the "AI 1.0 era". With the scars of Megvii's failed listing, SenseTime's continuously pressured stock price, Yitu's termination of the registration on the STAR Market, and Megvii's withdrawal of its IPO, at the age of 38, he played a new card at Step Star. In this company, there are two decision - makers: a Microsoft - affiliated scientist in his 50s as the CEO and a serial entrepreneur in his 38s as the chairman. This is the only dual - leadership structure among Chinese independent large model companies.
Tang Jie, Yan Junjie, Yang Zhilin, Jiang Daxin + Yin Qi. These are four different generations of people in the Chinese AI circle. They represent the old academic generation, the middle - aged generation from large companies, the young returnee generation, and the combination of serial entrepreneurs and technical experts respectively. The span of their birth years is from 1976 to 1992, a difference of 16 years. Their life trajectories hardly intersected until the spring of 2023, which Yang Zhilin called "a one - month window", when they each entered the same track with completely different judgments.
Three years later, in the Hong Kong stock market in May 2026, their respective judgments were translated into different numbers.
Who Saw Coding First
The biggest difference in judgment among these four people is not the technical route or the business model, but a simple thing: when did they each start to realize that Coding is the real main battlefield.
This timeline almost determined the market value rankings at the moment when the Hong Kong Stock Exchange closed on that afternoon in May 2026.
The first to see it was Tang Jie. Since Zhipu started betting on general large models in 2022, it has adhered to one thing: each generation of foundation models is made according to the standard of "comparable to the top - tier closed - source GPT models", rather than "making models according to what the product needs". In the context of the Chinese large model industry from 2023 - 2024, this judgment seemed out of place. Everyone was competing for users, entrances, and minds, while only Zhipu seemed to be "doing technology".
Looking back, this "out - of - place" is precisely the fundamental reason for Zhipu's highest market value in 2026. At the early morning of February 12, 2026, Zhipu released GLM - 5. With a total of 744B parameters and 40B active parameters in the MoE architecture, it scored 77.8 points in SWE - bench Verified, which was the highest score in the history of open - source models at that time (but later MiniMax M2.5's 80.2% has exceeded it). In the Artificial Analysis comprehensive intelligence index, GLM - 5 ranked first among global open - source models.
More importantly, GLM - 5.1, which was released one month later, according to the official report of Z.ai, scored 45.3 points in the Coding evaluation using Claude Code as the evaluation framework, reaching 94.6% of the score of Claude Opus 4.6. In April, GLM - 5.1 scored 58.4 points in SWE - Bench Pro, ranking first globally according to the list released by BenchLM, exceeding Claude Opus 4.6 and GPT - 5.4.
Behind these