GLM-4.5 Open Source Release: A New Generation of Flagship Models for Intelligent Body Applications

Smart Spectrum AI recently released its next generation of its flagship base model GLM-4.5, the model is designed for intelligent body (Agent) applications, and in the Hugging Face cap (a poem) ModelScope The platform is synchronized open source and its model weights follow the MIT license.

The model family utilizes a Mixed Expert (MoE) architecture and contains two versions:GLM-4.5 The total number of parameters is 355 billion and the activation parameter is 32 billion;GLM-4.5-Air The total number of participants is 106 billion with 12 billion activation parameters.MoE The architecture allows the model to activate only a portion of the expert network during inference, thus significantly reducing the actual computational overhead while maintaining the large knowledge scale, which is the key to the model's ability to achieve high efficiency.

In addition, the model offers two modes of operation: a "thinking mode" designed for complex reasoning and tool invocation, and a "non-thinking mode" designed for immediate response. On the cost side, the API is priced competitively, with high-speed versions generated at up to 100 percent faster. tokens/秒The

Overall performance

GLM-4.5 The goal is to natively fuse reasoning, coding, and intelligent body capabilities in a single model. To fully assess its generalized capabilities, the development team selected 12 industry-representative review benchmarks covering everything from expertise (MMLU Pro,AIME24), code generation (SWE-Bench Verified) to complex reasoning (GPQA) in multiple dimensions.

Figure : GLM-4.5 performance on 12 comprehensive benchmark reviews

In the combined average scores of these benchmark tests, theGLM-4.5 It is ranked third among models worldwide and first among open source models. According to official information, the model is at 15 trillion token After completing pre-training on generalized data in the domain of code, reasoning, and intelligences in the 8 trillion token The data was targeted for training, culminating in competency enhancement through intensive learning.

Higher parameter efficiency

In terms of parameter efficiency, theGLM-4.5 demonstrates the advantages of its architecture. Although its number of parameters is lower than DeepSeek-R1 cap (a poem) Kimi-K2 and other models, but performs better in several benchmarks.

Especially in measuring the ability of models to solve real software engineering problems SWE-Bench Verified Top of the list.GLM-4.5 The series is at the "Pareto frontier" of performance to parameter ratio. This means that the models in the series achieve the best performance available for the same parameter scale.

Figure: performance/parameter ratio of the model on the SWE-bench Verified list

Cost vs. speed

In addition to performance.GLM-4.5 The series is also a breakthrough in cost and efficiency. Its API call price is as low as $0.8/million input. tokensOutput $2/million tokensThis pricing is well below some of the leading models on the market. Meanwhile, high-speed versions are available up to 100 tokens/秒The generation speed is able to meet the actual deployment requirements of low latency and high concurrency.

Figure : Mainstream Model API Pricing Comparison

Real Scene Testing

In order to assess GLM-4.5 effect in real programming scenarios, the R&D team plugged it into the Claude Code framework, with Claude-4-Sonnet,Kimi-K2,Qwen3-Coder and other models were compared across 52 programming tasks covering six development domains.

Fig.: Comparison test results of real code smart body scenarios

The test results showed thatGLM-4.5 Compared to other open source models it excels in reliability of tool invocation and task completion, and can be used in most scenarios as a Claude-4-Sonnet of effective alternatives, but there is still room for improvement in the overall capability. To ensure the transparency of the evaluation, all test tasks and intelligent body trajectories have been made public.

Model Native Agent Scenarios

full-stack developer

GLM-4.5 The series is capable of performing full-stack development tasks, writing more complex applications, games and interactive web pages using natural language instructions. The development team demonstrated several examples of applications generated with just one command, which are now available in the Z.ai The website is online for users to experience for free.

Example 1: Building a Search Engine

Instructions: "Make a Google search site."
Experience Address: https://n0x9f6733jm1-deploy.space.z.ai

Example 2: Developing a Video Website

Instructions: "Develop a bilibili web demo with UI, pages include: home page and video details..."
Experience Address: https://n0dba6ce0e60-deploy.space.z.ai

Example 3: Developing a Social Media Site

Instructions: "Develop a Weibo web demo with UI, pages include: home and profile..."
Experience Address: https://v0rb06rruyf0-deploy.space.z.ai/

Artifacts effects

Models are not only good at code processing, but also at data processing and interactive content generation. For example, with an English command, a model can utilize the Three.js cap (a poem) JavaScript Create a 3D Visualize a globe, or make a Flappy Bird Style mini-games.

Example: Flappy Bird mini-game

Instructions: "Build a webpage using Three.js and JavaScript that creates a 3D world displaying places I've visited, based on an array. Clicking markers on the 3D globe will animate a zoom effect and open detailed trip information with photos."
Experience Address: https://chat.z.ai/space/b0yb2613ybp0-art

PPT production

GLM-4.5 It also demonstrated its ability in graphic content creation. Unlike traditional AI PPT tools that rely on templates to be filled in, the model autonomously searches for information, finds accompanying images, and generates content directly in HTML during the production process. This gives it the flexibility to create presentations, social media images or resumes at different scales.

How to use

GLM-4.5 Deeply optimized for full-stack programming and tool-calling capabilities, compatible with the Claude Code,Cline,Roo Code and other mainstream code intelligence body frameworks.

• Open Source Repository. https://github.com/zai-org/GLM-4.5
• Model Warehouse.
  • HuggingFace: https://huggingface.co/collections/zai-org/glm-45-687c621d34bda8c9e4bf503b
  • ModelScope: https://modelscope.cn/collections/GLM-45-b8693e2a08984f
• Online Experience.
  • HuggingFace: https://huggingface.co/spaces/zai-org/GLM-4.5-Space
  • ModelScope: https://modelscope.cn/studios/ZhipuAI/GLM-4.5-Demo