GLM-4.5V

// Overview

GLM-4.5V is the open-weight vision-language model in Z.ai's (Zhipu / GLM) GLM-V line, released on August 11, 2025. It is a Mixture-of-Experts model with 106B total parameters and 12B activated per token, built on top of the GLM-4.5-Air-Base text foundation. GLM-4.5V handles images (up to 4K resolution, arbitrary aspect ratio), video, multi-image prompts, and documents such as PDFs and slide decks, and it can output precise bounding-box grounding for elements in a scene.

Like the earlier GLM-4.1V-9B-Thinking, GLM-4.5V ships with a toggleable thinking mode: users can switch between fast direct answers and a deeper chain-of-thought pass for harder visual reasoning, OCR, chart and document parsing, video understanding, and GUI-agent tasks (screen reading, icon detection, desktop operation). Z.ai reports state-of-the-art results among similarly sized open models across 42 public vision-language benchmarks.

The weights are released under the MIT license on Hugging Face (zai-org/GLM-4.5V) and the GLM-V GitHub repo, allowing commercial use and fine-tuning. GLM-4.5V is also served through the Z.ai API platform and third parties such as OpenRouter, with an exposed context window of roughly 64K tokens and up to 16K tokens of output.

// Benchmarks

1. MMBench V1.1 (thinking)88.2%
2. MMStar (thinking)75.3%
3. MMMU val (thinking)75.4%
4. MMMU Pro (thinking)65.2%
5. MathVista (thinking)84.6%
6. AI2D (thinking)88.1%
7. OCRBench (thinking)86.5%
8. ChartQAPro (thinking)64%
9. ChartMuseum (thinking)55.3%
10. OSWorld (GUI agent, thinking)35.8%
11. WebVoyager (GUI agent, thinking)84.4%

Scores on a 0–100 scale (25-point gridlines); higher is better. Each benchmark links to its published source.

// Pricing

Z.ai API platform list price. Open weights (MIT) can also be self-hosted at no per-token cost.

Pricing source ↗

// Strengths

• Open weights under the permissive MIT license — free for commercial use and fine-tuning
• Efficient MoE design: 106B total parameters but only 12B activated per token
• Toggleable thinking mode balances fast responses against deeper multimodal reasoning
• Broad visual coverage: images up to 4K, video, multi-image, PDFs and slides in one pass
• Strong STEM and chart/document scores (MMMU, MathVista, AI2D, ChartQAPro)
• Native grounding with bounding-box output and GUI-agent control (OSWorld, WebVoyager)

// Best for

• Document, chart and long-PDF understanding and information extraction
• OCR and structured data extraction from images and plots
• Long-video segmentation and event recognition
• GUI agents: screen reading, icon detection and desktop operation assistance
• Visual grounding and spatial localization with bounding boxes
• Multi-image scene analysis, defect inspection and geo/context inference

// How to access

// GLM-V (vision-language) — every version

The full lineage of the GLM-V (vision-language) line, newest first. Every version has its own page — click any to compare specs, benchmarks and pricing.