Kimi broke through the open source ceiling, and global technology stocks evaporated US$470 billion on the 3 rd
21st Century Business Herald reporter Shi Enze In the early morning of July 17, Dark Side of the Moon announced the launch of KimiK3, the company’s most powerful large model to date. This model has become the world's largest open source model with 2.8 trillion parameters, and topped the authoritative front-end programming list FrontendCodeArena with 1679 points, surpassing Fable5 (1631 points) and GPT-5.6Sol (1618 points). This is the first time that an open source model has surpassed all overseas closed source models on the authoritative programming list.
21st Century Business Herald reporter Shi Enze In the early morning of July 17, Dark Side of the Moon announced the launch of Kimi K3, the company’s most powerful large model to date. This model has become the world's largest open source model with 2.8 trillion parameters, and topped the front-end programming authority list Frontend Code Arena with 1679 points, surpassing Fable 5 (1631 points) and GPT-5.6 Sol (1618 points). This is the first time that an open source model has surpassed all overseas closed source models on the authoritative programming list. The core reason why K3 topped the programming list and caused the U.S. stock market to plummet is that it has cut into the lifeblood of American technology giants: code generation is one of the largest tracks in the AI commercialization space, directly threatening the technical moat that overseas closed-source giants rely on to maintain high pricing. After the news was released, the Philadelphia Semiconductor Index fell 12.5% for the week, falling into a technical bear market; within 72 hours, the AI sector of the US stock market evaporated a total of approximately US$470 billion. Referring to the release of DeepSeek R1 last year, Nvidia's market value evaporated by US$589 billion in a single day. The market generally calls it "DeepSeek Moment 2.0". The impact is also directed at China's large model industry: on the day K3 was released, Hong Kong's "large model number one stock" Zhipu plummeted 28.49%, and MiniMax plummeted 15.62%. The competitive landscape and valuation logic of domestic large models changed accordingly. At the same time, the news of the launch of Dark Side of the Moon also emerged at the right time. On July 19, reporters learned from relevant sources that the company had sent a listing proposal to investors and expected to complete the Hong Kong stock listing within six months at the earliest. The company has not yet responded. From chip stocks to EDA, the AI industry chain has been comprehensively revalued Within 72 hours after the release of K3, the US stock AI industry chain suffered a comprehensive sell-off, and the pressure was gradually transmitted from the computing power layer to the application layer and tool layer. The computing power layer bears the brunt. Nvidia fell 5.3%, the Nasdaq 100 index fell 1.5%, and the Philadelphia Semiconductor Index fell 12.5% in a week, officially falling into a technical bear market. According to statistics, 17 Wall Street investment banks have lowered their target prices for AI chip companies overnight. JPMorgan Chase bluntly stated in the report that the release of Kimi K3 will undoubtedly make things worse. A Goldman Sachs partner also said that the era of computing power expansion may have come to an end. Technology giants at the model level are also under pressure. Meta fell more than 2.4%, and Google fell more than 2%. Market interpretation believes that the direct implication of K3 reaching the top is that the technical moat that the closed-source model relies on to maintain high pricing is being filled by open source. You must know that K3’s API pricing is only one-third of Fable 5 and 60% of GPT-5.6 Sol, and its performance has already surpassed programming in the core track of AI commercialization. The deepest decline was in the EDA sector, with Synopsys Technology closing down 7.85% and Cadence Electronics closing down 9.47%. Direct Fuse is an ability demonstrated by K3. K3 used open source EDA tools to independently complete the design, optimization and verification of a 4mm² chip in 48 hours, integrating 1.46 million standard units. An analyst who has been tracking the semiconductor industry for a long time said that when large models begin to complete chip design independently, the market is not only worried about the orders of one or two companies, but also the pricing logic of the entire traditional EDA tool chain. The shock wave was also transmitted to China’s AI peers in the Hong Kong stock market. On the day K3 was released, Hong Kong's "big model first stock" Wisdom (02513.HK) opened lower and closed lower, closing down 28.49% at HK$1,107. Its single-day market value evaporated by more than HK$200 billion, the largest single-day decline since its listing; MiniMax (00100.HK) closed down 15.62% at HK$216, with its total market value shrinking to HK$75.4 billion. The Hang Seng Technology Index fell 5% that day, and the AI sector of Hong Kong stocks came under collective pressure. In addition, SoftBank Group, which is regarded as a related target of OpenAI, also fell by about 9% in the Tokyo market.
Investment banks immediately caught up with this subtle change. Goldman Sachs pointed out in a research report released on July 18 that on the day K3 was released, market concerns about the competitive landscape of China's AI models were quickly reflected in the stock price, reflecting investors' high uncertainty about the long-term competitive landscape and sustainability of leadership positions of related companies. At the same time, Goldman Sachs emphasized that K3 set the API hybrid pricing at US$2.3 per million tokens, setting a new high in Chinese model pricing, marking that domestic AI companies are moving from a "price war" to a competition for "pricing power." However, some market analysts pointed out that it is unfair to attribute the plunge entirely to K3. Zhipu just experienced the lifting of the first batch of restricted shares on July 8, and then allotted nearly 20 million new H shares at a discount of HK$1,588 per share, raising approximately HK$31.4 billion. The market-to-sales ratio in 2025 once exceeded 1,000 times; MiniMax ushered in a large-scale lifting of more than 48% of its share capital on July 9, and its stock price has shrunk by about 80% from the historical high of HK$1,330 in March. This person believes that K3 is more like a fuse. The underlying reason is that the valuation of the AI sector is switching from "scarcity premium" to "fundamental pricing," and K3's peak has prompted the market to re-examine each company's technology moat. It is worth noting that this round of selling coincides with a sensitive window when the financing environment for the AI sector in the US stock market weakens. In the past few weeks, the investment-grade bond market has gradually struggled to digest a combined US$75 billion in debt issuance by Nvidia, SpaceX and Amazon. Dealogic data shows that the six AI giants (Alphabet, Amazon, Meta, Oracle, Nvidia and SpaceX) have issued approximately US$244 billion in bonds globally this year, far exceeding the US$108 billion in the whole of last year. Alphabet's 10-year bond spread widened by 12 basis points last week, and Meta's spread widened by 16 basis points, while the average investment-grade bond spread only edged up 2 basis points during the same period. "Everyone knows there are more bonds coming, so everyone is cautious about going all in at this point." Travis King, head of investment grade corporate bonds at Voya Investment Management, said. Industry insiders pointed out that K3's use of open source to achieve top performance has shaken the underlying logic of the "computing arms race" and amplified the market's existing concerns about over-investment in AI infrastructure. The dual reactions of stock prices and bond markets are essentially the same revaluation. Where does K3's lethality come from? In terms of hard indicators, K3 has a total parameter of 2.8 trillion, which is the world's first open source 3 trillion-level model, nearly double the 1.6 trillion of DeepSeek V4 Pro; the context window is 1 million tokens; it adopts the KDA hybrid linear attention architecture, activates 16 out of 896 experts, and supports native visual understanding. In addition, Dark Side of the Moon also announced that it will release the complete weight before July 27. In terms of running scores, K3 does not claim to completely surpass overseas closed-source models, but it ranks first in multiple productivity lists: SWE Marathon long-distance programming 42.0 points, Program Bench software reverse engineering 77.8 points, and Automation Bench office automation 30.8 points, all ranking first in the world; BrowseComp web page retrieval 91.2 points, basically the same as GPT-5.6 Sol; FrontierSWE difficult software engineering 81.2 points, second only to Fable 5. This is the first time that an open source model has surpassed all overseas closed source models on the authoritative programming list. However, the upper limit of K3’s capabilities coexists with the test of commercialization. The Dark Side of the Moon officially disclosed three limitations of K3: it is sensitive to historical thinking content, and switching models midway may lead to a decrease in output quality; training is biased towards long-distance and difficult tasks, and it may be overly proactive when faced with simple problems; it still lags behind Fable 5 and GPT-5.6 Sol in general reasoning and in-depth scientific research. According to actual test posts posted by many developers on social platforms, speed is K3’s current biggest shortcoming, and the generation time for the same scene is about two to three times that of GPT-5.6 Sol.
In terms of pricing, although K3 is lower than overseas closed-source models, it is 3 to 4 times more expensive than its previous generation K2.6. The output price has increased from 27 yuan/million tokens to 100 yuan/million tokens. Market acceptance still needs to be verified. Along with Kimi K3, there is also news that Kimi will be launched within 6 months. Public information shows that the company is valued at approximately US$31.5 billion (approximately 200 billion yuan), its ARR has exceeded US$300 million, and its cumulative financing in the three years since its establishment has exceeded 37 billion yuan. K3’s success in the core track of programming undoubtedly adds the heaviest weight to the listing proposal that has just been sent to investors. The Dark Side of the Moon accelerated its rush to Hong Kong stocks, which also added another layer of interpretation to the plunge of Zhipu and MiniMax on the day K3 was released. An analyst close to a Hong Kong stock investment bank told reporters that the heavy losses of the two companies were superficially caused by selling caused by the ebb and flow of model capabilities. The underlying reason was that funds were pricing in advance the changes in the pattern of the Hong Kong stock AI sector. "After the listing of Dark Side of the Moon, Hong Kong stocks will gather three leading domestic large-scale model companies. The scarcity premium that existing targets rely on to maintain high valuations will be further diluted, and funds will inevitably be reallocated between new stocks and old stocks." He further pointed out that, combined with the recent lifting and allotment pressures of Zhipu and MiniMax, the downward movement of the sector's valuation center may have just begun. In addition, compared with the shock caused by DeepSeek R1 last year, the target of this round of shock has shifted from "universal dialogue" to "code generation." The latter directly corresponds to developer productivity replacement and is regarded as one of the largest tracks in the AI commercialization space. The industry summarized the breakthrough moment of China's large models in one sentence: DeepSeek proved that "Chinese models can be cheap", the GLM series proved that "Chinese models can make money", and Kimi K3 tried to prove that Chinese models can "sit in the front row" in high-end productivity scenarios. However, reaching the top of a single list does not mean comprehensive surpassing. Overseas closed-source models still maintain the lead in general reasoning, in-depth scientific research and other dimensions, and the competition in computing power, capital and ecology is far from over. Whether K3 can transform "No. 1 on the list" into "No. 1 in market share" still needs to be observed in its developer retention and enterprise-level implementation progress after open source.
Investment banks immediately caught up with this subtle change. Goldman Sachs pointed out in a research report released on July 18 that on the day K3 was released, market concerns about the competitive landscape of China's AI models were quickly reflected in the stock price, reflecting investors' high uncertainty about the long-term competitive landscape and sustainability of leadership positions of related companies. At the same time, Goldman Sachs emphasized that K3 set the API hybrid pricing at US$2.3 per million tokens, setting a new high in Chinese model pricing, marking that domestic AI companies are moving from a "price war" to a competition for "pricing power." However, some market analysts pointed out that it is unfair to attribute the plunge entirely to K3. Zhipu just experienced the lifting of the first batch of restricted shares on July 8, and then allotted nearly 20 million new H shares at a discount of HK$1,588 per share, raising approximately HK$31.4 billion. The market-to-sales ratio in 2025 once exceeded 1,000 times; MiniMax ushered in a large-scale lifting of more than 48% of its share capital on July 9, and its stock price has shrunk by about 80% from the historical high of HK$1,330 in March. This person believes that K3 is more like a fuse. The underlying reason is that the valuation of the AI sector is switching from "scarcity premium" to "fundamental pricing," and K3's peak has prompted the market to re-examine each company's technology moat. It is worth noting that this round of selling coincides with a sensitive window when the financing environment for the AI sector in the US stock market weakens. In the past few weeks, the investment-grade bond market has gradually struggled to digest a combined US$75 billion in debt issuance by Nvidia, SpaceX and Amazon. Dealogic data shows that the six AI giants (Alphabet, Amazon, Meta, Oracle, Nvidia and SpaceX) have issued approximately US$244 billion in bonds globally this year, far exceeding the US$108 billion in the whole of last year. Alphabet's 10-year bond spread widened by 12 basis points last week, and Meta's spread widened by 16 basis points, while the average investment-grade bond spread only edged up 2 basis points during the same period. "Everyone knows there are more bonds coming, so everyone is cautious about going all in at this point." Travis King, head of investment grade corporate bonds at Voya Investment Management, said. Industry insiders pointed out that K3's use of open source to achieve top performance has shaken the underlying logic of the "computing arms race" and amplified the market's existing concerns about over-investment in AI infrastructure. The dual reactions of stock prices and bond markets are essentially the same revaluation. Where does K3's lethality come from? In terms of hard indicators, K3 has a total parameter of 2.8 trillion, which is the world's first open source 3 trillion-level model, nearly double the 1.6 trillion of DeepSeek V4 Pro; the context window is 1 million tokens; it adopts the KDA hybrid linear attention architecture, activates 16 out of 896 experts, and supports native visual understanding. In addition, Dark Side of the Moon also announced that it will release the complete weight before July 27. In terms of running scores, K3 does not claim to completely surpass overseas closed-source models, but it ranks first in multiple productivity lists: SWE Marathon long-distance programming 42.0 points, Program Bench software reverse engineering 77.8 points, and Automation Bench office automation 30.8 points, all ranking first in the world; BrowseComp web page retrieval 91.2 points, basically the same as GPT-5.6 Sol; FrontierSWE difficult software engineering 81.2 points, second only to Fable 5. This is the first time that an open source model has surpassed all overseas closed source models on the authoritative programming list. However, the upper limit of K3’s capabilities coexists with the test of commercialization. The Dark Side of the Moon officially disclosed three limitations of K3: it is sensitive to historical thinking content, and switching models midway may lead to a decrease in output quality; training is biased towards long-distance and difficult tasks, and it may be overly proactive when faced with simple problems; it still lags behind Fable 5 and GPT-5.6 Sol in general reasoning and in-depth scientific research. According to actual test posts posted by many developers on social platforms, speed is K3’s current biggest shortcoming, and the generation time for the same scene is about two to three times that of GPT-5.6 Sol.
In terms of pricing, although K3 is lower than overseas closed-source models, it is 3 to 4 times more expensive than its previous generation K2.6. The output price has increased from 27 yuan/million tokens to 100 yuan/million tokens. Market acceptance still needs to be verified. Along with Kimi K3, there is also news that Kimi will be launched within 6 months. Public information shows that the company is valued at approximately US$31.5 billion (approximately 200 billion yuan), its ARR has exceeded US$300 million, and its cumulative financing in the three years since its establishment has exceeded 37 billion yuan. K3’s success in the core track of programming undoubtedly adds the heaviest weight to the listing proposal that has just been sent to investors. The Dark Side of the Moon accelerated its rush to Hong Kong stocks, which also added another layer of interpretation to the plunge of Zhipu and MiniMax on the day K3 was released. An analyst close to a Hong Kong stock investment bank told reporters that the heavy losses of the two companies were superficially caused by selling caused by the ebb and flow of model capabilities. The underlying reason was that funds were pricing in advance the changes in the pattern of the Hong Kong stock AI sector. "After the listing of Dark Side of the Moon, Hong Kong stocks will gather three leading domestic large-scale model companies. The scarcity premium that existing targets rely on to maintain high valuations will be further diluted, and funds will inevitably be reallocated between new stocks and old stocks." He further pointed out that, combined with the recent lifting and allotment pressures of Zhipu and MiniMax, the downward movement of the sector's valuation center may have just begun. In addition, compared with the shock caused by DeepSeek R1 last year, the target of this round of shock has shifted from "universal dialogue" to "code generation." The latter directly corresponds to developer productivity replacement and is regarded as one of the largest tracks in the AI commercialization space. The industry summarized the breakthrough moment of China's large models in one sentence: DeepSeek proved that "Chinese models can be cheap", the GLM series proved that "Chinese models can make money", and Kimi K3 tried to prove that Chinese models can "sit in the front row" in high-end productivity scenarios. However, reaching the top of a single list does not mean comprehensive surpassing. Overseas closed-source models still maintain the lead in general reasoning, in-depth scientific research and other dimensions, and the competition in computing power, capital and ecology is far from over. Whether K3 can transform "No. 1 on the list" into "No. 1 in market share" still needs to be observed in its developer retention and enterprise-level implementation progress after open source.