GLM-4.5 - An open-source, state-of-the-art (SOTA) model for reasoning, code, and intelligent agents, developed by Zhipu.

What is GLM-4.5?

GLM-4.5 is Zhipu's next-generation flagship model, specifically designed for agent applications. It is the first open-source state-of-the-art (SOTA) model to natively integrate inference, code, and agent capabilities. Employing a hybrid expert (MoE) architecture, it comes in two versions: GLM-4.5 (355 billion parameters, 32 billion activations) and GLM-4.5-Air (106 billion parameters, 12 billion activations). The model demonstrates outstanding performance across multiple benchmarks, achieving top-tier overall performance among open-source models, particularly excelling in code-based agent scenarios. It supports a hybrid inference mode, providing both thinking and non-thinking modes to cater to both complex tasks and immediate response requirements.

Parameter efficiency is doubled, API price is only 1/10 of Claude's, and speed can reach up to 100 tokens/second. It is the first to achieve native fusion of multiple capabilities in a single model and has achieved a major technological breakthrough - a single model has powerful inference, code, and intelligent agent capabilities at the same time, and is available for free trial on Zhipu Qingyan and Z.ai.

Key features of GLM-4.5

• Multi-capability integrationFor the first time, it has achieved a native fusion of reasoning, code generation, and agent capabilities in a single model, which can meet the needs of complex agent applications.
• reasoning abilityIt performs exceptionally well across multiple inference benchmarks, supports complex inference tasks, and achieves top-tier inference performance compared to open-source models.
• Code generationIt excels in code generation and programming tasks, supports multiple programming languages, generates high-quality code snippets, and is capable of handling full-stack development tasks.
• Intelligent agent applicationsIt supports functions such as tool invocation and web browsing, and can be integrated with code intelligence frameworks such as Claude Code and Roo Code, making it suitable for intelligence tasks.
• Hybrid reasoning modeIt offers a "thinking mode" for complex reasoning and tool use, and a "non-thinking mode" for immediate response, balancing efficiency and performance.

Technical Highlights of GLM-4.5

• Hybrid Expert (MoE) ArchitectureGLM-4.5 employs a Mixture of Experts (MoE) architecture, a highly efficient model expansion method. The MoE architecture divides the model parameters into multiple expert modules and dynamically selects some experts for activation during each forward propagation, achieving sparse activation of the parameters. This significantly reduces computational cost and memory usage while maintaining the model's powerful expressive capabilities.
  • GLM-4.5The total number of parameters is 355 billionActivation parameters are 32 billion.
  • GLM-4.5-AirThe total number of parameters is 106 billionActivation parameters are 12 billion.
• Multimodal capabilitiesGLM-4.5 supports multimodal input and output, and can handle various data types such as text and images. This enables the model to better understand and generate complex agent application content, such as handling multimodal interaction scenarios like web browsing and tool invocation in agent tasks.
• Hybrid reasoning modeGLM-4.5 offers two inference modes:
  • Thinking ModeSuitable for complex reasoning and tool usage scenarios, the model will engage in deeper thinking and reasoning to generate more accurate results.
  • Non-Thinking ModeSuitable for real-time response scenarios, the model generates results quickly, meeting the requirements of low latency.
• Efficient training processThe training process for GLM-4.5 consists of three stages:
  • General Data PretrainingPre-training on general data of 15 trillion tokens to learn the basic representations of language and knowledge.
  • Targeted trainingTargeted training is conducted on 8 trillion tokens of data in fields such as code, reasoning, and intelligent agents to enhance the model's performance on specific tasks.
  • Reinforcement learning optimizationFurther optimization of the model's reasoning, code generation, and agent capabilities through reinforcement learning can improve the model's performance in real-world applications.
• Parameter efficiency optimizationGLM-4.5 achieves a significant improvement in parameter efficiency. Despite having only half the number of parameters as DeepSeek-R1 and one-third as many as Kimi-K2, it outperforms many standard benchmarks. For example, in a model family with 100B total parameters – 10B activation parameters, GLM-4.5's inference performance on the Artificial Analysis benchmark rivals or even surpasses other top international models.

GLM-4.5 project address

• GitHub repository: https://github.com/zai-org/GLM-4.5
• HuggingFace warehousehttps://huggingface.co/collections/zai-org/glm-45-687c621d34bda8c9e4bf503b
• ModelScope repository: https://modelscope.cn/collections/GLM-45-b8693e2a08984f
• Experience address:
  • HuggingFacehttps://huggingface.co/spaces/zai-org/GLM-4.5-Space
  • ModelScope: https://modelscope.cn/studios/ZhipuAI/GLM-4.5-Demo

Technical Specifications of GLM-4.5

• comprehensivePerformance SOTA:The 12 most representative benchmarks include MMLU Pro, AIME 24, MATH 500, SciCode, GPQA, HLE, LiveCodeBench, SWE-Bench Verified, Terminal-Bench, TAU-Bench, BFCL v3, and BrowseComp.comprehensiveAverage scoreGLM-4.5 achievedGlobal ModelNo.threeDomesticModelNo.oneopen sourceModelFirst.

• Higher parameter efficiencyThe GLM-4.5 has half the number of parameters as DeepSeek-R1 and one-third as many as Kimi-K2, yet it outperforms in multiple standard benchmarks, thanks to the higher parameter efficiency of the GLM model. On the SWE-bench Verified leaderboard, which measures model code capability, the GLM-4.5 series is at the Pareto front in terms of performance/parameter ratio, indicating that it achieves the best performance at the same scale.

• Low cost, high speedBeyond performance optimization, the GLM-4.5 series also achieves breakthroughs in cost and efficiency, resulting in pricing significantly lower than mainstream models: API call prices are as low as 0.8 yuan per million tokens for input and 2 yuan per million tokens for output. Meanwhile, the high-speed version has been tested to generate up to 100 tokens per second, supporting real-world deployment needs for low latency and high concurrency, balancing cost-effectiveness and user experience.

How to use GLM-4.5

• Experience Platform
  • Zhipu Qingyan (chatglm.cn)Users can directly access the Zhipu Qingyan platform to experience the full functionality of GLM-4.5 for free. Users can perform various operations such as dialogue generation, code generation, and inference tasks to experience the model's powerful capabilities.
  • Z.aiUsers can also experience the features of GLM-4.5 through the Z.ai platform.
• API calls
  • BigModel.cnZhipu AI provides an API interface, which users can call through the BigModel platform. The API interface supports various functions, including text generation, code generation, and inference tasks.

GLM-4.5 model pricing

API call prices are as low as: 0.8 yuan per million tokens for input and 2 yuan per million tokens for output, with a generation speed of up to 100 tokens per second, supporting low latency and high concurrency deployment.

Application scenarios of GLM-4.5

• Full-stack development taskGLM-4.5 is capable of handling complex full-stack development tasks and supports the development of relatively complex applications, games, interactive web pages, etc.
• Code generationGLM-4.5 excels in code generation, producing high-quality code snippets and supporting multiple programming languages. It can help developers quickly generate code frameworks, fix code errors, and optimize code structure.
• Programming aidsModels can serve as programming aids, providing features such as code completion, code generation suggestions, and code comments to improve development efficiency.
• Content generationThe model can generate various types of content, such as articles, news reports, and creative copywriting, and is suitable for content creation, copywriting, and other scenarios.
• academic researchGLM-4.5 can be used for academic research, helping researchers explore cutting-edge issues in fields such as natural language processing and artificial intelligence.