GLM-4.5V: Z.ai's GLM-4.5 Vision Model, 64K Context

What GLM-4.5V is

GLM-4.5V is the vision-language model of Z.ai's GLM-4.5 generation, MIT-licensed and self-hostable. Its context window is 64K tokens, recorded here as 65.5K, which is half what its text-only stablemates GLM-4.5 and GLM-4.5-Air carry and a third of what the GLM-4.7 line offers. Images consume context, so the window fills faster with image input.

LMArena

On LMArena's text leaderboard GLM 4.5V has 1352 Elo (rank 218); by category, 1402 Elo in Coding (rank 212), 1360 Elo in Math (rank 214), 1399 Elo in Expert (rank 184) and 1309 Elo in Creative Writing (rank 227). LMArena's text board presents written prompts only; it does not show the model an image, so it does not measure visual performance.

Successor

GLM-4.6V is the next vision model in the same line, MIT-licensed on identical terms, with a 128K window. Z.ai have since published GLM-5V Turbo, which is proprietary rather than MIT.

What it is built on

Z.ai built GLM-4.5V on GLM-4.5-Air, with 106 billion total parameters and 12 billion active, continuing the approach of GLM-4.1V-Thinking. They report state-of-the-art results among models of the same scale on 42 public vision-language benchmarks. The card lists image reasoning including multi-image analysis and spatial recognition, long-video segmentation and event recognition, GUI tasks such as screen reading, icon recognition and desktop operation, chart and long-document parsing, and visual grounding. It has the same thinking-mode switch as the GLM-4.5 language model. GLM-4.5V was released on 11 August 2025.

Grounding output and deployment

For grounding, the model returns bounding boxes as four coordinates, top-left and bottom-right, normalised to a 0 to 1000 scale relative to the image, marked with the special tokens <|begin_of_box|> and <|end_of_box|>. Z.ai publish a desktop assistant app that connects to GLM-4.5V and captures screenshots or screen recordings, and example prompt templates for GUI agents on mobile, PC and web. The model runs on Transformers 4.57.1 or later, vLLM 0.10.2 or later and SGLang 0.5.3 or later, with a tensor-parallel size of four in Z.ai's serving examples. Its predecessor GLM-4.1V-9B-Thinking supports any aspect ratio and images up to 4K resolution.

On the Z.ai API

Z.ai's API documentation lists GLM-4.5V at $0.60 per million input tokens and $1.80 output, with video, image, text and file input, text output and a maximum output of 16K tokens. It describes the thinking-mode switch as letting users move between quick responses and deep reasoning, and lists seven use cases: web page coding, grounding, GUI agent tasks, complex document interpretation, image recognition and reasoning, video understanding, and subject problem solving for education. Z.ai state performance at the level of open-source state-of-the-art models across 42 public visual multimodal benchmarks.

The GLM-V line

GLM-4.5V belongs to Z.ai's GLM-V series, described in the paper on GLM-4.1V-Thinking and GLM-4.5V. GLM-4.1V-9B-Thinking, released on 1 July 2025 on the GLM-4-9B-0414 foundation, introduced reinforcement learning with curriculum sampling and a 64K context, and Z.ai open-sourced the reward system used to train it. GLM-4.5V and GLM-4.1V share multimodal preprocessing but use different conversation templates, so prompts must match the model.

Official Resources