On August 6th, major Hong Kong stock indices experienced a collective correction, and Asia-Pacific technology stocks were generally under pressure. However, concept stocks related to large models in Hong Kong stocks moved independently against the trend: MiniMax surged by over 22% intraday, and Zhipu rose by over 10%. Two major catalysts resonated on the same day: the adjustment of Stock Connect targets officially took effect today, with both MiniMax and Zhipu gaining "entry tickets" from mainland capital. Coupled with the open-sourcing of MiniMax's H3 multimodal model, which triggered a collective adaptation trend among domestic chips, the investment logic for the large model sector is being revalued.
Large Model Giants Lead Against the Trend with "Living Water" Injected by Stock Connect
On the day's trading, Hong Kong stocks opened lower and declined. The Hang Seng Index closed down 1.49%, and the Hang Seng Tech Index fell 2.28%. Memory concept stocks led the decline, while chip stocks and large-cap internet stocks generally fell. SMIC fell nearly 4%, Baidu Group fell by over 4%, and Alibaba and Xiaomi Group fell by nearly 3%. Amid the overall sluggish atmosphere, the two large model leaders became the focus of the market: MiniMax surged by 22.77% intraday, reaching a high of HKD 311.60 per share, and maintained a gain of over 17% by the close, with a market capitalization exceeding HKD 100 billion; Zhipu rose by 10% intraday, reaching a high of HKD 1145 per share, and its market capitalization surpassed HKD 500 billion.
The direct catalyst for the market came from significant changes in the trading system. On the morning of August 6th, the Shenzhen Stock Exchange announced that the target securities list for Stock Connect was adjusted to include MINIMAX-W, Innolux, and Sanhuan Group, effective from that day. On the evening of August 5th, the Shanghai Stock Exchange also announced that the aforementioned targets were added to the Stock Connect list for Hong Kong Stock Connect. This means that starting today, mainland investors can directly buy and sell these two AI large model leaders through the Stock Connect channel. The expectation of "living water" injection from Southbound capital provided direct liquidity support for the stock price fluctuations. It is worth mentioning that Zhipu was the first to be included in the connection on June 8th, and MiniMax officially "entered the connection" after meeting the stable trading period. The "twin stars" of domestic large models have thus gathered in the Stock Connect channel.
H3 Open-Sourcing Combined with Domestic Chip Adaptation Pushes Software and Hardware Synergy into Substantive Implementation Stage
Beyond the market movements, the fundamental catalysts are also worth noting. On August 3rd, MiniMax announced the official open-sourcing of its general-purpose multimodal generative system, H3. According to public information, H3 can uniformly understand multimodal contexts composed of text, images, videos, and audio, and can generate videos up to 2K resolution, 15 seconds long, with native stereo audio. It possesses extensive multimodal context understanding and generation capabilities during the pre-training phase and can excellently follow complex multimodal instructions, making it one of the leading multimodal systems in the current global open-source community.
What excited the market even more was the speed of the industry chain's response—on the day H3 was open-sourced, multiple domestic chip manufacturers, including GigaDevice, Moore Threads, Biren Technology, and Muxi, simultaneously announced the completion of adaptation. Some brokerage firms pointed out that the collective completion of adaptation by leading domestic chip manufacturers on the first day of model open-sourcing both verified the friendliness of the H3 architecture to domestic hardware and signified that the software and hardware synergistic ecosystem of "domestic models + domestic computing power" has entered a substantive implementation stage. This is expected to further reduce enterprise deployment costs, expand application scope, and provide continuous traction for upstream computing infrastructure. Capital-level movements also confirm the industry's fervor. Information from Qichacha shows that Shanghai Xiyu Jizhi Technology Co., Ltd., an affiliated company of MiniMax, increased its registered capital from 4 billion RMB to 5.5 billion RMB recently, with two major capital injections this year totaling 4.5 billion RMB, demonstrating the determination of leading players to invest in the AI large model track. Furthermore, international institutions such as Jefferies and Goldman Sachs have recently issued buy ratings, with institutional calculations suggesting potential incremental funds exceeding 10 billion HKD after the IPO.
Domestic large models undergo intensive iteration, and the Chinese contingent dominates the global usage ranking.
Looking at the bigger picture, the strength of domestic large model concepts is not an isolated event. Recently, the public beta of the official DeepSeek-V4-Flash API was released, Moonshot AI launched and open-sourced Kimi K3 with a 2.8 trillion parameter MoE architecture, and Zhipu GLM series continues to iterate... Domestic large models are continuously approaching top international levels in advanced versions, while general-purpose versions are expanding their usage scale based on the open-source ecosystem and high cost-effectiveness.
The latest data from OpenRouter, a global multi-model aggregation platform, shows that the top six large language models in the global usage ranking in July are all open-source large models independently developed in China, including Xiaomi MiMo-V2.5, DeepSeek-V4-Flash, Tencent Hy3, MiniMax M3, Zhipu GLM-5.2, and DeepSeek-V4-Pro. Behind the "domestic contingent's dominance" is a narrative shift in the Chinese AI industry from "catching up" to "running alongside," which has also become an important anchor for the capital market to re-evaluate the valuation of AI concept stocks. According to disclosures from the National Development and Reform Commission, as of the end of June, the national intelligent computing power reached 2.8 times that of the same period last year, the total global downloads of domestic large models exceeded 10 billion times, and a virtuous cycle of mutual promotion among computing power, models, and data is accelerating.
Institutional analysis: Both computing power and application are key tracks, and commercialization realization is crucial.
Amid the market excitement, institutions remind investors to pay more attention to the rhythm of fundamental realization. Bohai Securities believes that domestic large models, relying on open-source strategies, are expected to participate in the global open-source market competition, accelerating their international expansion, while also attracting some overseas enterprise clients to switch from closed-source models to open-source solutions. Meanwhile, United Securities points out that domestic large models are forming a combined advantage of "high-performance open-source models + low-cost inference services + multimodal scenario capabilities," and AI office is expected to become the next certain direction for token consumption after code generation, with enterprise-level intelligent agent applications accelerating implementation.
From the industry chain transmission, computing power demand is spreading from the "training end" to the "massive inference end." Recently, several listed companies have disclosed large computing power orders: a subsidiary of Fengzhushou signed a computing power service contract worth 4.608 billion RMB, a subsidiary of Gaole shares signed a computing power service contract worth 3.195 billion RMB, and Xingyun Technology has a long-term framework order for computing power worth 15.082 billion RMB. A research report from Bank of China International suggests that the decline in token prices does not necessarily suppress the total computing power demand, and the higher token consumption brought by multi-round planning and tool calling in Agents is expected to boost inference computing power demand, continuously strengthening the positive logic of the domestic computing power industry chain. However, market risks cannot be overlooked. On August 6th, the quarterly performance guidance released by the overseas storage leader, Samsung, fell short of expectations, once again raising concerns in the market about the vitality of AI hardware. The Korean composite index briefly dropped by over 5% and triggered a "circuit breaker," putting pressure on Asia-Pacific technology stocks. Analysts point out that after the significant rebound in Hong Kong stocks, the performance of technology stocks in the second and third quarters remains a key factor determining the overall market direction. Investors need to closely monitor global AI capital expenditure trends and changes in the Federal Reserve's policy, guard against high-level volatility risks, and seize structural opportunities during sector rotation. (The content of this article is compiled based on public information and is for reference only, and does not constitute investment advice. Markets are risky, and investment requires caution.)