A newly launched AI mannequin from China is drawing widespread reward from builders, producing a stage of pleasure and hype not seen for the reason that emergence of DeepSeek greater than a 12 months in the past.
Chinese language AI startup Zhipu AI launched its newest giant language mannequin (LLM) referred to as GLM-5.2 final week, designed to run long-coding duties and agentic workflows. Providing a a million token context window, GLM-5.2 is taken into account to be in the identical league as frontier fashions developed by US rivals comparable to Anthropic’s Claude Opus 4.8 and OpenAI’s GPT 5.5.
Just like DeepSeek’s fashions, GLM-5.2 is an open-weight mannequin. Which means that anybody can obtain the weights of the LLM and host it regionally on their very own techniques whereas making adjustments to it as they see match. In distinction, most American frontier fashions like these from Anthropic and OpenAI are closed.
Since its debut, GLM-5.2 has been producing quite a lot of intrigue throughout social media networks with traders, founders, builders, and others expressing awe on the high-performance and capabilities of the brand new mannequin. “Genuinely impressed, nearly shocked, at how good GLM-5.2 by @zai_org is at coding. This adjustments issues,” Guillermo Rauch, CEO of Vercel, a cloud-based platform for builders, wrote on X.
After spending a complete day utilizing GLM-5.2, Matt Velloso, a former vice chairman of Meta, Google DeepMind, and Microsoft, mentioned, “First open mannequin that passes the bar as a every day driver. Issues will not be going to be the identical.” In a current change on X, tech trillionaire Elon Musk estimated {that a} Chinese language AI mannequin able to matching Anthropic’s Claude Fable 5 would probably arrive within the first quarter of subsequent 12 months. Tang Jie, co-founder and chief scientist at Zhipu replied that it “gained’t take that lengthy”.
Investor optimism fueled by the thrill round 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 growing as a lot as 42 per cent to a peak of HK$2,980, in line with a report by South China Morning Submit.
However is the hype round GLM-5.2 justified? What units the most recent Chinese language open-weight AI mannequin aside? How does it evaluate with rival fashions from OpenAI, Anthropic, and Google? And extra importantly, will its launch have implications for the rising US-China battle for AI supremacy? Let’s have a look.
Story continues under this advert
What’s Zhipu AI?
Zhipu was based in 2019 by Tang Jie and Li Juanzi, who had been professors within the Laptop Science Division at Tsinghua College in Beijing. The startup is strongly backed by the Chinese language authorities, with Zhipu thought-about to be part of China’s ‘AI tigers’, which is a gaggle of promising AI startups within the nation vying with the US to steer the event of this frontier know-how.
Zhipu went public on the Hong Kong Inventory Trade in January 2026, alongside rival MiniMax, one other AI tiger. Zhipu trades underneath the title Data Atlas Know-how. Liu Debing, chairman of Data Atlas Know-how, has amassed a fortune of $22.4 billion for the reason that itemizing, in line with a report by Forbes.
In February 2026, the corporate unveiled an AI mannequin referred to as GLM-5 that might purpose and write code. On the time, Zhipu mentioned the mannequin’s efficiency was similar to Claude Opus 4.5, a product launched by Anthropic final November. The corporate launched GLM-5.1, an up to date model of the mannequin, in April this 12 months. Zhipu claimed that it had made breakthroughs in response pace in addition to different enhancements with GLM-5.1.
In its current observe, JPMorgan raised its 2026-2030 income forecast for Zhipu by between 7 and 16 per cent following the launch of GLM-5.2. Stating that the mannequin strengthened Zhipu’s management in China’s AI sector and confirmed the corporate’s rising pricing energy, the financial institution projected an over 534 per cent income surge for Zhipu in 2026, and anticipated the AI agency to show a revenue in 2028, reversing its earlier forecast of a internet loss for that 12 months.
Story continues under this advert
Beneath the hood of GLM-5.2
GLM 5.2 is a 753 billion-parameter AI mannequin whose core weights have been launched underneath an unrestricted MIT open-source license.
In a technical paper launched alongside the mannequin, the researchers mentioned that they achieved a significant architectural optimization referred to as ‘IndexShare’ with GLM-5.2. In normal LLMs, recalculating consideration mechanisms throughout lengthy paperwork takes up quite a lot of compute. Nonetheless, IndexShare solves this by reusing the an identical indexer throughout each 4 sparse consideration layers. In consequence, it reduces per-token compute FLOPs by 2.9 instances.
The mannequin additionally boasts an upgraded Multi-Token Prediction (MTP) layer for speculative decoding, which boosts accepted token size by as much as 20 per cent throughout inference. As well as, GLM-5.2 comes with completely different ‘Considering Modes’ comparable to ‘Max’ or ‘Excessive’ that customers can choose to regulate the mannequin’s reasoning abilities.
Benchmark efficiency
When evaluated on industry-standard third-party benchmark exams, GLM-5.2 achieved scores larger than most open-weight flagship fashions, together with DeepSeek v4. Its efficiency was additionally above or on par with OpenAI’s GPT-5.5 and Anthropic’s Claude Opus 4.8.
Story continues under this advert
It at present ranks second globally on the Code Area listing for front-end internet growth duties, trailing solely Anthropic’s flagship Claude Fable 5. GLM-5.2 even beat out Claude Fable 5 on a crowdsourced design process benchmark referred to as Design Area, with an ELO rating of 1360.
The mannequin is especially good at agentic device use and long-horizon software program engineering duties, scoring 62.1 on SWE-bench Professional which is larger than GPT-5.5 (58.6). When examined for long-horizon process completion, GLM-5.2 scored 74.4 per cent on the FrontierSWE (Dominance) benchmark surpassing GPT-5.5 (72.6 per cent) and ending in a near-tie with Claude Opus 4.8 (75.1 per cent).