Prateek Tanmay
A newly released AI model from China is drawing widespread praise from developers, generating a level of excitement and hype not seen since the emergence of DeepSeek more than a year ago.
Chinese AI startup Zhipu AI launched its latest large language model (LLM) called GLM-5.2 last week, designed to run long-coding tasks and agentic workflows. Offering a one million token context window, GLM-5.2 is considered to be in the same league as frontier models developed by US rivals such as Anthropic’s Claude Opus 4.8 and OpenAI’s GPT 5.5.
Similar to DeepSeek’s models, GLM-5.2 is an open-weight model. This means that anyone can download the weights of the LLM and host it locally on their own systems while making changes to it as they see fit. In contrast, most American frontier models like those from Anthropic and OpenAI are closed.
Since its debut, GLM-5.2 has been generating a lot of intrigue across social media networks with investors, founders, developers, and others expressing awe at the high-performance and capabilities of the new model. “Genuinely impressed, almost shocked, at how good GLM-5.2 by @zai_org is at coding. This changes things,” Guillermo Rauch, CEO of Vercel, a cloud-based platform for developers, wrote on X.
After spending an entire day using GLM-5.2, Matt Velloso, a former vice president of Meta, Google DeepMind, and Microsoft, said, “First open model that passes the bar as a daily driver. Things are not going to be the same.” In a recent exchange on X, tech trillionaire Elon Musk estimated that a Chinese AI model capable of matching Anthropic’s Claude Fable 5 would likely arrive in the first quarter of next year. Tang Jie, co-founder and chief scientist at Zhipu replied that it “won’t take that long”.
Investor optimism fueled by the buzz around GLM-5.2 led to the Hong Kong-listed AI startup to surpass HK$1 trillion (US$128 billion) on Monday morning, June 22, with shares increasing as much as 42 per cent to a peak of HK$2,980, according to a report by South China Morning Post.
But is the hype around GLM-5.2 justified? What sets the latest Chinese open-weight AI model apart? How does it compare with rival models from OpenAI, Anthropic, and Google? And more importantly, will its launch have implications for the growing US-China battle for AI supremacy? Let’s take a look.
Story continues below this ad
What is Zhipu AI?
Zhipu was founded in 2019 by Tang Jie and Li Juanzi, who were professors in the Computer Science Department at Tsinghua University in Beijing. The startup is strongly backed by the Chinese government, with Zhipu considered to be a part of China’s ‘AI tigers’, which is a group of promising AI startups in the country vying with the US to lead the development of this frontier technology.
Zhipu went public on the Hong Kong Stock Exchange in January 2026, alongside rival MiniMax, another AI tiger. Zhipu trades under the name Knowledge Atlas Technology. Liu Debing, chairman of Knowledge Atlas Technology, has amassed a fortune of $22.4 billion since the listing, according to a report by Forbes.
In February 2026, the company unveiled an AI model called GLM-5 that could reason and write code. At the time, Zhipu said the model’s performance was comparable to Claude Opus 4.5, a product released by Anthropic last November. The company launched GLM-5.1, an updated version of the model, in April this year. Zhipu claimed that it had made breakthroughs in response speed as well as other improvements with GLM-5.1.
In its recent note, JPMorgan raised its 2026-2030 revenue forecast for Zhipu by between 7 and 16 per cent following the launch of GLM-5.2. Stating that the model reinforced Zhipu’s leadership in China’s AI sector and showed the company’s growing pricing power, the bank projected an over 534 per cent revenue surge for Zhipu in 2026, and expected the AI firm to turn a profit in 2028, reversing its previous forecast of a net loss for that year.
Story continues below this ad
Under the hood of GLM-5.2
GLM 5.2 is a 753 billion-parameter AI model whose core weights have been released under an unrestricted MIT open-source license.
In a technical paper released alongside the model, the researchers said that they achieved a major architectural optimization called ‘IndexShare’ with GLM-5.2. In standard LLMs, recalculating attention mechanisms across long documents takes up a lot of compute. However, IndexShare solves this by reusing the identical indexer across every four sparse attention layers. As a result, it reduces per-token compute FLOPs by 2.9 times.
The model also boasts an upgraded Multi-Token Prediction (MTP) layer for speculative decoding, which boosts accepted token length by up to 20 per cent during inference. In addition, GLM-5.2 comes with different ‘Thinking Modes’ such as ‘Max’ or ‘High’ that users can select to adjust the model’s reasoning skills.
Benchmark performance
When evaluated on industry-standard third-party benchmark tests, GLM-5.2 achieved scores higher than most open-weight flagship models, including DeepSeek v4. Its performance was also above or on par with OpenAI’s GPT-5.5 and Anthropic’s Claude Opus 4.8.
Story continues below this ad
It currently ranks second globally on the Code Arena list for front-end web development tasks, trailing only Anthropic’s flagship Claude Fable 5. GLM-5.2 even beat out Claude Fable 5 on a crowdsourced design task benchmark called Design Arena, with an ELO score of 1360.
The model is particularly good at agentic tool use and long-horizon software engineering tasks, scoring 62.1 on SWE-bench Pro which is higher than GPT-5.5 (58.6). When tested for long-horizon task completion, GLM-5.2 scored 74.4 per cent on the FrontierSWE (Dominance) benchmark surpassing GPT-5.5 (72.6 per cent) and finishing in a near-tie with Claude Opus 4.8 (75.1 per cent).
