Zhipu suddenly plummeted over 20%! The big bullish stock is rapidly diving! What happened?

Big stocks, significant pullback.

On the 23rd, in the Hong Kong stock market, “the world’s leading AI model stock” Zhipu’s share price reversed the trend and plummeted, with intraday declines exceeding 25%. By the close of trading, Zhipu’s decline remained at 22.76%.

Market analysts pointed out that Zhipu and MINIMAX had recently experienced huge gains, leading many funds to take profits. Additionally, an apology letter released by Zhipu on the 21st may have also put pressure on the company’s stock price on the 23rd.

Zhipu’s stock price drops over 20%

On the 23rd, Hong Kong-listed AI large model concept stocks saw a sharp decline. Intraday, Zhipu fell over 25%, and MINIMAX dropped more than 15%. By the close, Zhipu and MINIMAX had fallen 22.76% and 13.35%, respectively.

In terms of news, on the evening of February 21st, Zhipu issued an apology regarding the GLM Coding Plan, stating that after the release of GLM-5, three mistakes were made: insufficient transparency of rules, slow rollout of the latest large model GLM-5’s grayscale process, and rough upgrade mechanisms for old users.

Zhipu explained that after the release of GLM-5, traffic exceeded expectations, and the company’s expansion pace could not keep up. As a result, GLM-5 was gradually opened in the order of Max, Pro, and Lite. Currently, Max users are fully open, Pro users are also open but may experience rate limiting during peak times due to high cluster load, and Lite users will be gradually opened in grayscale after the holiday during non-peak periods. The company supports affected Lite and Pro users (regardless of new or old) to apply for refunds independently.

Public information shows that Zhipu was founded in 2019, originating from the knowledge engineering laboratory of the Department of Computer Science at Tsinghua University. It possesses full-stack independent R&D capabilities from underlying algorithms and pre-training frameworks to domestic hardware adaptation for large models. The company is committed to using large models to achieve or surpass human capabilities in language, reasoning, vision, hearing, and tool use. It is reported that Zhipu developed the fully domestic pre-training architecture GLM based on autoregressive fill-in, achieving the best results in robustness, controllability, and hallucination, with models adapted to over 40 domestic chips.

By 2025, Zhipu plans to launch a new generation flagship model, GLM-4.5/4.6, which will be the first to natively integrate reasoning, encoding, and agent capabilities in a single model. The model immediately ranked first domestically and first globally among open-source models in 12 authoritative evaluations. On the global large model marketplace OpenRouter (reflecting actual global large model usage), GLM-4.5/4.6 has maintained a top 10 call volume since launch, and paid API revenue has surpassed that of all other domestic models combined.

Currently, the GLM series large models are applied in industries such as public governance, industrial manufacturing, energy and electricity, finance, internet, communications, consumer electronics, and education, serving over 2.7 million customers and developers.

Stock price once surged over 500%

Before the sharp decline on the 23rd, Zhipu’s stock price had been continuously rising. On the previous trading day (February 20th), the first Hong Kong stock trading day after the Year of the Horse Spring Festival, Zhipu’s stock surged over 40% against the trend, with a market value surpassing HKD 320 billion.

When it listed on January 8th this year, Zhipu’s issue price was HKD 116.2, and its closing price on February 20th was HKD 725. In just over a month, the company’s stock price skyrocketed by 524%.

On February 12th, Zhipu officially launched and open-sourced its new flagship model GLM-5. That day, Zhipu’s stock price jumped 28.68%, and in the following days, it continued to rise.

Zhipu explained that academia and industry are gradually forming a consensus that large models are evolving from coding and front-end development to engineering and completing large tasks, transforming from “Vibe Coding” to “Agentic Engineering.” GLM-5 is a product of this transformation: achieving state-of-the-art open-source performance in coding and agent capabilities, with user experience approaching Claude Opus 4.5 in real programming scenarios, excelling in complex system engineering and long-term agent tasks. In the authoritative Artificial Analysis ranking, GLM-5 ranks fourth globally and first among open-source models.

Ten days later, on February 22nd, the technical report for GLM-5 was released. Zhipu stated that GLM-5 is a next-generation foundational model aimed at promoting the shift from “Vibe Coding” to “Agentic Engineering” in programming paradigms. Building on the previous GLM-4.5, GLM-5 enhances intelligent agent, reasoning, and coding (ARC) capabilities by adopting sparse attention (DeepSeek Sparse Attention, DSA) to significantly reduce inference costs while maintaining long context capabilities.

Zhipu added that to better align the model with various tasks, they built a new asynchronous reinforcement learning (RL) infrastructure, decoupling generation and training processes to greatly improve post-training iteration efficiency. They also proposed a new asynchronous agent reinforcement learning algorithm to further enhance reinforcement learning effects, enabling the model to learn more effectively from complex, long-term interactions. Thanks to these innovations, GLM-5 achieved state-of-the-art performance on mainstream open benchmarks. Most importantly, GLM-5 demonstrated unprecedented capabilities in real-world programming tasks, surpassing all previous open-source baselines in handling end-to-end software engineering challenges.

Changjiang Securities pointed out that since the beginning of this year, leading domestic model companies Zhipu and MINIMAX have successively listed on the Hong Kong Stock Exchange on January 8th and 9th, both performing well in the market. Combining stock price reviews, with the release of their new generation foundational models in February, stock prices continued to hit new highs. The core reason is that as model capabilities improve, domestic models are entering the demand era—advancing from capability enhancement to scene capability expansion, demand growth, high-quality revenue streams, and a positive data feedback loop that further boosts model capabilities.

What expectations are reflected behind the stock price? What investment opportunities are worth noting? Changjiang Securities states that two major changes mark a key turning point: first, domestic models are once again catching up with the currently most practical overseas Claude Opus series; second, improvements in model capabilities will gradually open up revenue potential. Future pricing will mainly consider market share. Currently, domestic models in the coding field have become a highly cost-effective global choice. As domestic market capacity increases, overseas market share growth will present huge opportunities for domestic model manufacturers. Leading investment opportunities include new super portals, domestic foundational resources, and AI agents.