Zhipu Trillion Market Value Singularity: Is GLM-5.2 really far ahead? How to translate Over the past week, the Hong Kong stock market has pressed the fast forward button on Zhipu ⏩⏩⏩ On June 15th, Zhipu rose nearly 48% during trading and closed up 32.82%; On June 18th, the stock price rose another 26.14%, and the market value surged to approximately HKD 933.6 billion; On June 22nd, Zhipu reached a high of HKD 2980 during trading, with a total market value exceeding HKD 1 trillion. In just 5 trading days, Zhipu's market value has risen by one Meituan, three MINIMAX, and four JD.com This is undoubtedly a carnival between holders and the Hong Kong stock market, but we have seen too many scenes in the technology industry where gods are created and unfinished. When promoting GLM-5.2 as the "Deepseek Moment 2.0" of domestic large-scale models, it is worth pondering how much of the soaring stock price of Zhipu is real money and how much is the emotional premium of a scarce AI asset? one ️⃣ Why was Zhipu suddenly sold out? Zhipu, a domestic large model company with a total revenue of 724 million yuan in 2025 and a loss of 4.718 billion yuan for the whole year, has suddenly been pushed onto the trillion yuan market value table by the market. This round of market is not an ordinary good brought by the release of a single model, but multiple narratives collided at the same time. The first is the availability event of Anthropic Fable 5/Mythos 5. On June 12th, Anthropic issued a statement stating that due to US government export control directives, all customer access to Fable 5 and Mythos 5 must be suspended. The White House's geopolitical policy on AI tools has forced global developers to think about a very practical question: no matter how strong the AI model is, what if I can't use it myself? Secondly, the release nodes of GLM-5.2 are too precise. After the ban was issued, Zhipu quickly launched GLM-5.2 and announced that the model has a 1M long context, is oriented towards long-range coding and Agentic Engineering, and follows the MIT protocol for open source. When the US closed source frontier model began to experience access uncertainty, Zhipu presented the market with an "open, cheap, and deployable" alternative card. Thirdly, there is a scarcity of first-hand AI targets in China. In fact, Zhipu is one of the few pure large-scale model targets that can be directly priced by the Chinese capital market. Deepseek previously raised more than 50 billion yuan, with an estimated value of about 50 billion dollars. However, it was only open to state-owned capital or Internet giants such as the National Artificial Intelligence Industry Investment Fund, Tencent, Ningde Times, and all capital shares were locked up for five years, so it was impossible to cash out in a short time. And other domestic big models are concentrated in the hands of comprehensive technology giants such as ByteDance, Alibaba, and Tencent. In contrast, the stock price of Zhipu is more like a domestic big model Beta. As long as the product is successful and the narrative is established, it is easier for funds to concentrate and flow in. Fourthly, there is an expectation of going public on the A-share market. Zhipu has advanced the relevant procedures for listing on the Science and Technology Innovation Board and has previously disclosed plans to raise RMB 15 billion. For the market, this means that it is not just a Hong Kong AI concept stock, but may also become a domestic AI infrastructure company under the A+H dual capital platform. So this round of rise is not just about the market value discovery of Zhipu's GLM-5.2 model, but also about the market realizing that in the context of unstable global AI access, domestically produced open-source leading models may have new strategic premiums. two ️⃣ What is the strength of GLM-5.2? If we only look at marketing rhetoric, GLM-5.2 seems to have already punched Claude and kicked GPT, dominating the world and being self-centered. But a more accurate statement is that GLM-5.2 did not crush Claude/GPT in all dimensions, but achieved catch-up in long-range coding, agentic engineering, and open weighting. According to the official model card of Zhipu, GLM-5.2 is a MoE model with 744B total parameters and 40B activation parameters, supporting 1M token context. It introduces the IndexShare architecture, reduces per token FLOPs by 2.9 times under 1M context length, and optimizes MTP speculative decoding to increase inference acceptance length. In the official benchmark, GLM-5.2 shows significant improvement compared to GLM-5.1: Terminal-Bench 2.1：81.0 vs 63.5 SWE-bench Pro：62.1 vs 58.4 FrontierSWE：74.4 vs 30.5 PostTrainBench：34.3 vs 20.1 ➡️ These indicators all point in the same direction: the application core of GLM-5.2 is long-term, multi-step, cross file, and tool calling engineering tasks. Third party reviews also gave GLM-5.2 a high rating. Artificial Analysis ranks GLM-5.2 as the top open source weight model with an Intelligence Index score of 51, surpassing open-source or open weight models such as MiniMax-M3, DeepSeek V4 Pro, and Kimi K2.6. In the Code Arena WebDev ranking, GLM-5.2 ranks second, only behind Claude Fable 5, and scores higher than models such as Claude Opus 4.8 Thinking. More importantly, the ranking price of GLM-5.2 is $1.4 per million token for input and $4.4 per million token for output, while Fable 5 is $10 per 50. In summary, from the perspective of the ability spectrum, GLM-5.2 is probably in an interesting position: It is not the strongest model, but it is already the state-of-the-art (State of the Art) of current open source models. It is not the cheapest model, but it is already the most cost-effective among the same level of ability. It is not the most versatile model, but it is unique in long-range engineering and agent tasks. three ️⃣ Can GLM-5.2 support the trillion dollar market value of Zhipu? Artificial Analysis also pointed out that GLM-5.2 outputs an average of 43k tokens per Intelligence Index task, of which 37k are reasoning tokens, which is higher than multiple open-source models of the same level. This indicates that GLM-5.2 is strong, but also very token burning. In real commercialization, eating tokens means cost, latency, and gross profit margin pressure. The ultimate goal for big model companies is not just ranking scores, but whether each call can generate real revenue. The real silver side is that in 2025, Zhipu's total revenue will be 724 million yuan, a year-on-year increase of 131.9%; The ARR of MaaS API platform is about 1.7 billion yuan, an increase of 60 times in the past year; Http://(BigModel. cn) has connected more than 4 million enterprises and developers, and nine of China's top ten Internet companies have made deep calls to GLM. However, the foam on the other side is also obvious: in 2025, Zhipu will lose 4.718 billion yuan in the year, with an adjusted net loss of 3.182 billion yuan and R&D expenditure of about 3.18 billion yuan. That is to say, Zhipu is still in the stage of "exchanging huge R&D investment for model capabilities and developer ecology". According to its $1 billion ARR guidance by the end of 2026, Zhipu's P/ARR was about 94 times higher at that time, while Anthropic was about 18 times higher. The release of GLM-5.2 has indeed given Zhipu a strong presence in the global developer context for the first time, which is very important for a large Chinese model company. But Zhipu still needs to prove three things to confirm its identity as a "Chinese Anthropic": Firstly, can the developer popularity of GLM-5.2 be converted into continuous API calls and paid retention. Secondly, can long-range coding and agent capabilities truly enter the production process of enterprises. Thirdly, can the strategic influence of open source models in turn enhance the commercial capture capability of Zhipu. However, as of the time of publication, the stock price of Zhipu has rebounded to HKD 2470 within the day The editor wants to ask, is there a god who rushes in around 3000 Hong Kong dollars?