Introduction

Chinese AI company Zhipu has officially announced the release of its flagship model, GLM-4.5, on the evening of the 28th. This model is designed specifically for intelligent agent applications, achieving breakthroughs in complex reasoning, code generation, and agent interaction. It has ranked among the top globally in comprehensive testing performance. According to Techi, the development of AI in China is showing a clear trend of accessibility, as Zhipu not only competes with industry giants like DeepSeek and OpenAI but also strives to stand on the same stage with them.

Performance Highlights

The key performance points of GLM-4.5 include achieving open-source SOTA (state-of-the-art) capabilities in reasoning, coding, and agent functions. The development team evaluated the model’s general capabilities using 12 representative benchmarks, where GLM-4.5 ranked third globally, first among domestic models, and first among open-source models. Users can access this model through the Zhipu open model platform, with its open-source version also available on Hugging Face and ModelScope.

This new model signifies an important shift in AI development. It aims to evolve from being a passive question-answering chatbot to an AI employee capable of understanding complex goals and autonomously planning and executing multi-step tasks. For instance, users can instruct GLM-4.5 to independently develop a search-enabled website similar to Google, create a Bilibili-like platform with bullet comments, or even launch a complete Flappy Bird game with a simple command.

Training Process

GLM-4.5 and GLM-4.5-Air utilize a similar training process. Initially, they were pre-trained on 15 trillion tokens of general data, followed by targeted training on 80 trillion tokens in areas such as coding, reasoning, and agents. Finally, reinforcement learning was applied to enhance the model’s reasoning, coding, and agent capabilities. GLM-4.5 has half the parameters of DeepSeek-R1 and one-third of Kimi-K2, yet it performs better in several standard benchmark tests due to its higher parameter efficiency.

Within just 10 hours of its release, GLM-4.5 attracted global media attention. Reports highlighted its lower costs and superior performance, with some media providing detailed analyses and comparisons of its parameters. Currently, the API call price for this series of models is as low as 0.8 yuan per million tokens for input and 2 yuan per million tokens for output, with the high-speed version reaching up to 100 tokens per second.

Global Reception

Less than two hours after the model’s release, it was featured on the homepage of X platform. Within 12 hours, it climbed to second place on the global Hugging Face leaderboard, setting a record for growth rate. CNBC noted that the AI models being developed by Chinese companies are not only becoming smarter but also more cost-effective, echoing the core breakthrough advantages that DeepSeek had when it shook the market.

Indian financial media MINT commented that the launch of GLM-4.5 coincides with a period of increasing activity among Chinese AI startups, which are striving to enhance their market visibility and industry recognition in a field long dominated by American companies. The continuous release of more powerful and accessible models may inject new momentum into domestic and international AI competition.

Reuters reported on the 28th that Zhipu, one of China’s “AI Six Tigers,” has gained attention since being named by OpenAI in June. Bloomberg also mentioned that Zhipu AI, a challenger to OpenAI, is attempting to seize the high ground in this rapidly developing industry by establishing industry standards for future technological advancements. VentureBeat highlighted that the release of GLM-4.5 provides a viable, high-performance foundational model for teams balancing innovation, performance, and operational constraints.

Open Source Strategy

Another highlight of the GLM-4.5 model is its full open-source release under the MIT license. Zhipu CEO Zhang Peng explained that Zhipu is one of the earliest AI companies in China to open-source large models, having done so multiple times. In 2022, they open-sourced a hundred billion parameter base model, which attracted widespread attention globally, prompting major companies like Apple and OpenAI to download and analyze it. Earlier this year, when ChatGPT gained popularity, Zhipu released a 6B model that could run directly on laptops. This has led to increased public understanding of open-source strategies.

Zhang emphasized that open-sourcing at this juncture is a significant strategic direction for both academia and national strategy. He believes that this move can encourage the industry and academia to rethink the development of AI and explore the prospects of AGI (Artificial General Intelligence), potentially leading to a new computational paradigm.

The Next Paradigm

Some Western media outlets have attempted to label Zhipu as “China’s OpenAI.” Zhang noted that while both companies began training large models early in their respective countries, their paths and philosophies differ. Both companies have been exploring various directions to push the boundaries of AI capabilities.

Zhang pointed out that existing models are not yet true general models, as some excel in programming, others in mathematics, and some in reasoning, but none perform optimally across all tasks. GLM-4.5 aims to integrate these diverse capabilities into a single model, striving to become a “perfect student” with increasingly general abilities.

Zhang believes that the global AI leaders may be reaching a consensus that foundational models are crucial. He stated that Zhipu does not adopt a fully closed approach like OpenAI but instead fosters scientific development and technological advancement through openness and collaboration in academia and industry.

Exploring AI Limits

In light of the recent U.S. announcement of the “AI Action Plan,” which some have dubbed the “AI Star Wars,” Zhang discussed its potential impact on the AI competition landscape between the U.S. and China and global AI development. He noted that the U.S. strategy of reinforcing its advantages in AI may lead to a new level of competition. However, the essence of AI development remains the exploration of the limits of AGI capabilities, and it is challenging to predict the costs involved in this endeavor.

Zhang emphasized the importance of recognizing the direction of AGI technology development and maintaining focus on the future rather than just immediate concerns. He also highlighted the need to explore the limits of intelligence and identify genuine application scenarios for AI.

Zhipu’s vision for the future involves enhancing understanding of AGI technology trends and exploring the upper limits of AI capabilities, while also ensuring that their development path aligns with the needs of China’s AI growth.