zai-org/glm-4.7

GLM-4.7 is a flagship text model launched by Zhipu, primarily positioned to “strengthen agentic coding capabilities, covering complex task solving and efficient creation across multiple scenarios.” It can autonomously plan, execute, and verify code tasks.

• Key Upgrades: Compared to GLM-4.6, its mathematical and reasoning abilities have improved by 41%, achieving a HLE benchmark score of 42.8%, surpassing GPT-5.1; it introduces retained thinking and round-based thinking modes, ensuring more stable execution of complex tasks and allowing on-demand control over reasoning overhead.
• Applicable Scenarios: It covers various professional scenarios, including agentic coding, multimodal interaction app development, frontend design, office productivity, in-depth research, software development, algorithm implementation, and script automation.
• Evaluation Data: The τ²-Bench interactive tool invocation evaluation scored 84.7 points (open-source SOTA); the BrowseComp web task evaluation scored 67 points.
• Product Value: It generates complete, runnable code frameworks, reducing manual assembly and debugging costs and accelerating the process from idea to implementation.
• Core Parameters: Context window size: 200K; maximum output tokens: up to 128K; supports structured format outputs such as JSON.
• Evaluation Data: It supports over 100 programming languages, with an average of 3.2 toolchain calls per task, and an autonomous repair rate exceeding 65%.

───────────────────────────────────────────────────────────────────

Core Capabilities

💻 Strong Breakthrough in Agentic Coding: From requirement understanding to solution decomposition, it autonomously completes integration across multiple tech stacks, generating structurally complete, runnable code that covers front-end and back-end integration scenarios.

🧠 Controllable Thinking Modes: Supports interleaved, retained, and round-based thinking. For simple tasks, it reduces thinking latency; for complex tasks, it increases accuracy by enabling full-fledged thinking; and for long-term tasks, it provides more efficient caching.

🎨 Enhanced Frontend Aesthetic Quality: It precisely understands UI design specifications and offers high-quality default solutions for layout, color schemes, and component styles, significantly reducing time spent on fine-tuning styles.

📝 Immersive Creation Experience: Text expressions are delicate and vividly visual, character roles remain consistent with established personas, plot progression feels natural, and office creations adhere closely to mainstream formatting standards.

───────────────────────────────────────────────────────────────────

Test Data