zai-org/glm-4.5v

GLM-4.5V is a flagship multimodal visual reasoning model launched by Zhipu AI, with a core focus on setting a new open-source benchmark for “all-scenario visual understanding + deep reasoning”.

• Ultra-large-scale MoE architecture: Total parameters: 10.6B; active parameters: only 1.2B; achieves state-of-the-art (SOTA) performance among open-source models at the same level in 41 public multimodal benchmarks.
• Full-modal input support: Natively processes images, videos, PDFs/Word documents, and GUI screen captures, enabling unified visual reasoning.
• Outstanding real-world scenario capabilities: Accurately locates object coordinates, replicates web frontend code, infers the latitude and longitude of street-view shooting locations, and even ranks among the top 66 globally in human competitions.
• Flexible switching between thinking modes: Supports both “fast response” and “deep reasoning” modes, striking a balance between efficiency and output quality.

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

Core Capabilities

👁️ Precise Visual Grounding: Outputs precise bounding boxes based on natural language descriptions (e.g., “the second beer bottle from the right”), suitable for industrial applications such as quality inspection and remote sensing.

🖥️ GUI Intelligent Agent Operation: Understands e-commerce pages and office software interfaces, automatically recognizes prices and icons, and executes actions such as clicks and edits.

📄 Deep Interpretation of Complex Documents: Reads and understands images and text like a human, simultaneously comprehending charts, tables, and text without losing information through OCR.

🎬 Long-Video Semantic Reasoning: Analyzes dynamic changes across multiple video frames, reconstructs interaction logic, and even replicates runnable HTML pages from screen recordings.

🧠 World Knowledge Fusion Reasoning: Combines visual clues such as vegetation, buildings, and climate to infer image shooting locations and background information without requiring external searches.