AI sectors in A-shares and Hong Kong stocks collectively surged.
On April 9, the AI large model sector in Hong Kong stocks continued its strong performance, with Xunce Technology (03317.HK) and Zhipu (02513.HK) both hitting record highs since listing during the session. By the close, Xunce Technology was at 288 HKD per share, up 24.03% from the previous trading day, with an intraday high of 289.4 HKD per share and a market cap reaching 92.94 billion HKD; Zhipu was at 933 HKD per share, up 8.74%, with an intraday high of 998.5 HKD per share and a market cap once breaking through 430 billion HKD. On April 10, Xunce Technology and Zhipu continued to rise at the open, with Xunce Technology up over 10% at the time of writing.
The A-share AI sector led the surge. On April 8, the China Merchants创业板 AI ETF (159243.SZ) surged 10.05% throughout the day with heavy volume, with component stocks like BlueFocus (300058.SZ) and Yidian Tianxia (301171.SZ) leading the gains, reflecting high market sentiment. On April 9, BlueFocus continued to rise, closing at 16.58 CNY per share, up 1.66% from the previous day; the benchmark index for the China Merchants创业板 AI ETF rose 9.47% that day.
In terms of news, a text-to-video model, HappyHorse-1.0, topped the authoritative evaluation platform Artificial Analysis榜单, outperforming ByteDance's Seedance 2.0 in scores; Zhipu launched its flagship model GLM-5.1, capable of working for 8 hours continuously. Domestic large models demonstrated unexpectedly strong technical capabilities, while commercialization步伐 significantly accelerated. Multiple catalytic factors combined to ignite a full-line rally in the capital market, from computing infrastructure to AI application terminals.
Regarding the recent sector-wide rise, Time Weekly reporters sent letters to Yidian Tianxia and called BlueFocus for comments but received no response by the time of writing.
Xiang Anling, an associate professor at the School of Journalism and Communication at Minzu University of China, told Time Weekly that the current iteration节奏 of AI models has明显 accelerated, with almost daily new feature releases and new models topping evaluation lists periodically, which includes some market speculation. However, she noted that the recently released GLM-5.1 and HappyHorse-1.0 indeed show noteworthy亮点 in model capabilities. Xiang Anling's research focuses on AIGC and media big data, and she leads the National Natural Science Foundation project "AIGC Risk Identification."
Why Did Two Large Models Drive Market Sentiment?
"Now, one key indicator of AI strength is whether it requires human intervention or can work independently," said Xiang Anling.
She believes that the biggest breakthrough of Zhipu's GLM-5.1 is its ability to work like a real software engineer for 8 hours continuously, autonomously planning, executing, testing, correcting errors, and delivering complete engineering results, with almost no human intervention needed throughout the process.
Previous large models were more like "temporary workers," answering user queries one by one or stopping after writing a piece of code to wait for human review. But GLM-5.1 is different—it can understand complex full tasks, arrange what to do over the next 8 hours, proactively change plans when encountering bottlenecks, and fix errors on its own.
In terms of programming capability, in the SWE-bench Pro benchmark test, which simulates real software development scenarios, GLM-5.1 scored 58.4, surpassing Claude Opus 4.6 (57.3 points) and GPT-5.4 (57.7 points), marking the first time a domestic open-source model has outperformed top overseas closed-source models on this metric. More importantly, this model was entirely trained on Huawei's Ascend 910B chips, without using any NVIDIA GPUs, which is a significant milestone in China's AI autonomy journey. The market sees not just the progress of one model but the validation of a complete closed loop of "domestic computing power + domestic model."
Another dark horse also performed impressively. HappyHorse-1.0 emerged anonymously in early April 2026 and topped the authoritative AI evaluation platform Artificial Analysis's Video Arena list on the night of April 7. In the text-to-video (without audio) category, its Elo score (a comprehensive ranking score derived from repeated "head-to-head" matches,直观 reflecting model strength in real user preferences) soared to 1357 points, leading Seedance 2.0 by 84 points.
HappyHorse-1.0 can complete text-to-video generation in one go, with synchronized video and audio output. Traditional AI video generation mostly produces silent footage, requiring separate audio processing that is hard to match accurately. HappyHorse-1.0 can automatically add sounds like ice cracking or basketball swishing based on scene descriptions and supports lip-sync for seven languages.
In terms of generation efficiency, HappyHorse-1.0 uses a lightweight design with only 15 billion parameters, far fewer than most competitors. With DMD-2 distillation technology, HappyHorse-1.0 takes about 38 seconds to generate a 1080p高清 video and only 2 seconds for a low-resolution preview.
Regarding its幕后 team, though官方 initially did not respond, multiple media outlets reported that the team behind it is Alibaba's Taotian Group's Future Living Lab, led by Zhang Di, the "father of Kling," who developed this product in just 5 months after returning to Alibaba from Kuaishou. Affected by this news, Alibaba's Hong Kong stock (09988.HK) surged over 7% directly on the afternoon of April 7.
Time Weekly reporters sought confirmation from Alibaba but received no response by the time of writing.
Xiang Anling stated that there might be differences between榜单成绩 and actual落地 tasks, requiring more scenario testing for verification. Additionally, with the current rapid iteration speed, new models may soon surpass existing achievements, so continuous practice-based testing and observation are still needed.
AI Has Passed the Market Education Phase
Whether the AI industry has entered a new cycle of "order落地 and profit release" is being increasingly affirmed by the market.
While releasing GLM-5.1, Zhipu announced a 10% price increase for its API服务, with pricing for coding scenarios matching that of Anthropic's Claude Sonnet 4.6. This is the first time a domestic large model has achieved price parity with overseas leading vendors in core scenarios.
Regarding the price hike, Zhipu told Time Weekly that longer推理链路, increased token consumption, and larger models have raised推理 costs, and the price increase is to restore the model's normal commercial value.
Behind this is a change in industry logic—Zhipu is no longer sacrificing profits for market share but pricing based on cost and value.
In the "2026 Global AI Commercial落地 Value Insight Research Report,"亿欧智库 proposed that the global AI industry in 2026 is shifting from "technology competition" to "value co-creation," moving from the scale logic of "stacking computing power and parameters" to the efficiency logic of "intensive cultivation and profit creation." The report introduces the VPT (value per token) evaluation system, emphasizing that enterprises need to increase the ratio of economic value to token consumption to achieve profitability.
In other words, industry consensus is forming: AI cannot remain in the money-burning validation phase forever; it must move towards sustainable business models.
Xiang Anling's judgment is relatively cautious but clear: although the industry is only in its early stages, it is no longer in the state of purely burning money for market validation and free promotion as last year or before. Now, especially some leading or oligopolistic AI models, have begun规模化 value monetization. She gave two personal examples: after Seedance 2.0 was released, as a paid member, she had to wait four or five hours in queue to generate a video; and for Zhipu's programming model, she has been trying to buy a service package since before the New Year but still can't get it even by logging in at 10 a.m. every day. These experiences直观 indicate that high-quality model services are already in short supply, users are willing to pay, and even after paying, they have to queue.
This means that leading models are starting to gain real paying users through technical barriers and scarcity. Vendors are no longer shy about raising prices but openly adjust them based on cost and commercial value. Although this is still far from a full-blown profit harvesting period, with many mid-to-long-tail models still seeking monetization paths and computing costs remaining high, the industry inflection point is clear: the AI industry has indeed moved beyond the purely free 'market education phase' and entered the early stage of 'order落地 and profit release'.
This article is from WeChat public account "Time Weekly" (ID: timeweekly), author: Li Jiaxuan, editor: Wang Ying.
