AB
AB
AiBoss

All Categories

project

GLM-4.7 - Zhipu's flagship AI model, further enhancing its coding capabilities.

GLM-4.7 is an open-source model launched by Zhipu AI, focusing on improving coding capabilities, inference abilities, and tool collaboration. The model excels in multi-language programming, complex task planning, and front-end design aesthetics, and supports various programming frameworks...

What is GLM-4.7?

GLM-4.7 is an open-source model launched by ZAIPA, focusing on improving coding, inference, and tool collaboration capabilities. The model excels in multi-language programming, complex task planning, and front-end design aesthetics, supporting multiple programming frameworks such as Claude Code. In benchmark tests, GLM-4.7's coding capabilities reach leading levels among open-source models, with significantly enhanced inference abilities. The model introduces interleaved, retained, and wheel-based thinking modes, making complex task execution more stable and controllable. The model is now available as an API service through BigModel and has been launched in the Skills module within z.ai's full-stack development model, providing developers with an efficient and intelligent programming experience.

Main functions of GLM-4.7

  • Core coding capabilitiesGLM-4.7 excels in multi-language programming and terminal tasks, supports a "think before you act" approach, and significantly improves the stability and code quality of complex tasks.
  • Front-end design optimizationIt supports the generation of more modern and beautiful web pages and slideshows, improving UI design quality and reducing the time developers spend on style adjustments.
  • Tool calling capabilityIt performs exceptionally well in tool calls and web browsing tasks, achieving a τ²-Bench score of 87.4% and a BrowseComp score of 67.5, demonstrating significant improvements in efficiency and accuracy.
  • Complex reasoning abilityThe computer's mathematical and reasoning abilities have been significantly enhanced, with an HLE benchmark score of 42.8%, an improvement of 12.4% over the previous generation, enabling it to handle complex logic and mathematical problems.
  • Thinking pattern optimizationIt introduces staggered, retention, and round-robin thinking modes to improve the stability and controllability of complex tasks, making it suitable for long-term tasks and multi-round dialogues.
  • Multimodal and full-stack developmentIt supports multimodal task collaboration and full-stack development, integrates the Skills module, and helps developers build applications with rich interactions and smooth user experience.

GLM-4.7 performance

  • Core Coding Capabilities:
    • SWE-bench VerifiedThe score was 73.8%, an improvement of 5.8 percentage points compared to GLM-4.6, reaching the state-of-the-art (SOTA) level of open-source models.
    • SWE-bench MultilingualThe score was 66.7%, an improvement of 12.9 percentage points from GLM-4.6, indicating a significant enhancement in multilingual programming ability.
    • Terminal Bench 2.0The score was 41%, an improvement of 16.5 percentage points compared to GLM-4.6, indicating a significant improvement in terminal task performance.
  • Tool Using Capability:
    • τ²-BenchThe score was 87.4%, an improvement of 12.2 percentage points compared to GLM-4.6, and the interactive tool calling capability reached the open source state-of-the-art level.
    • BrowseComp (web browsing)The score was 52.0%, an improvement of 6.9 percentage points compared to GLM-4.6; in the BrowseComp test with context management enabled, the score was 67.5%, an improvement of 10.0 percentage points compared to GLM-4.6, showing better performance in web browsing and toolchain management.
  • Complex Reasoning:
    • HLE (Humanity's Final Exam)The score was 42.8%, an improvement of 12.4 percentage points from GLM-4.6, indicating a significant enhancement in mathematical and reasoning abilities.
    • MMLU-ProThe score was 84.3%, an improvement of 1.1 percentage points from GLM-4.6, demonstrating stable performance in multi-domain reasoning ability.
    • GPQA-DiamondThe score was 85.7%, an improvement of 4.7 percentage points from GLM-4.6, indicating a further improvement in reasoning accuracy.

Project address for GLM-4.7

  • Project official website: https://z.ai/blog/glm-4.7
  • GitHub repository: https://github.com/zai-org/GLM-4.5
  • HuggingFace model libraryhttps://huggingface.co/zai-org/GLM-4.7

Application scenarios of GLM-4.7

  • Programming and Software DevelopmentGLM-4.7 can generate high-quality multilingual code, serving as an intelligent programming assistant to improve development efficiency.
  • Front-end development and designIn web design and UI/UX design, it can quickly generate modern and beautiful layouts and color schemes, reducing the time front-end developers and designers spend adjusting styles.
  • Complex Task Planning and ExecutionWith its reserving and wheel-based thinking modes, GLM-4.7 can handle complex multi-step tasks, ensuring accuracy and stability for long-term missions.
  • Education and LearningGLM-4.7 provides code examples and exercises for programming education, while also helping students improve their thinking skills through math and logic problem training.
  • Office AutomationGLM-4.7 can automatically generate documents, reports, and data analysis code, reducing manual writing and formatting time and improving office efficiency.
Back