Z.ai, previously known as Zhipu, has unveiled its next-generation base model, GLM-4.5. This innovative release marks a significant milestone, achieving top-tier performance among open-source models and redefining what’s possible in artificial intelligence, making powerful capabilities more accessible and affordable for developers.
GLM-4.5 represents a major leap, being Z.ai’s first entry into the open-source world with a “Mixture of Experts” (MoE) architecture. This advanced system is available in two powerful versions: the flagship GLM-4.5, boasting 355 billion total parameters, and the more streamlined GLM-4.5-Air, with 106 billion total parameters.
The GLM-4.5 series stands out by seamlessly combining reasoning, coding, and agent-like abilities within a single model, designed to handle complex automated tasks.
The model’s performance has been rigorously tested across 12 key benchmarks, covering its reasoning, coding, and agent capabilities. GLM-4.5 secured an impressive third place globally based on average scores, while also ranking first among all domestic and open-source models. GLM-4.5-Air demonstrates exceptional efficiency, making it a leading model within its category of roughly 100 billion parameters.
Also Read: Samsung secures $16.5B deal to build Tesla’s next-generation AI chips
Beyond raw power, the GLM-4.5 series also offers notable advantages in processing speed and cost. Developers can access the model with API calls priced as low as USD 0.11 per million input units and USD 0.28 per million output units. The high-speed version can generate content at a rate exceeding 100 units per second, enabling rapid development.
GLM-4.5 is unique as Z.ai’s first foundational model with built-in “Agent-native” capabilities. This means it naturally understands how to reason and act, allowing it to autonomously plan multi-step processes, create detailed data visuals, and manage entire workflows from start to finish. This integration of diverse abilities makes GLM-4.5 highly suitable for sophisticated applications that require combined intelligence, perception, and action.
Unlike many proprietary systems, GLM-4.5 will be released under an open, verifiable license, offering businesses the flexibility of on-premise deployment and customization. This approach provides greater control and transparency, positioning GLM-4.5 as a truly open alternative in an industry often dominated by closed systems. Z.ai CEO Zhang Peng highlighted this commitment, stating, “We are setting a new benchmark with GLM-4.5, demonstrating that cutting-edge performance can be open, efficient, and affordable.” He added that GLM-4.5 is their first complete realization of integrating more general intelligent capabilities without losing existing ones.
Z.ai is increasingly recognized as a globally competitive AI company from China, with OpenAI itself identifying it as one of the few global firms capable of developing comparable models at competitive prices. The company’s standing was further acknowledged in Stanford University’s Artificial Intelligence Index Report 2025, where it was included for the first time among organizations developing “notable AI models.” Demonstrating its commitment to responsible development, Z.ai was also the first among its Chinese counterparts to sign the Frontier AI Safety Commitments.
Also Read: While AI erased 80,000 tech jobs, it created huge salary boosts elsewhere
Since 2020, Z.ai’s open-source models have garnered over 40 million downloads worldwide. Its diverse portfolio includes the versatile GLM Series, from the large 355 billion parameter model down to 6 billion, supporting multiple languages. Their AutoGLM model, China’s earliest agent model, has even surpassed more recent offerings from OpenAI and Anthropic in agent-related assessments.
The company also offers an Agent Ecosystem of tools for building custom AI assistants and “Flash Models” that achieve high performance with a smaller footprint, allowing for faster and more cost-effective deployments.
