GLM-4.5V

What it is

GLM-4.5V is a vision-language model — one that takes images, video and text as input and answers in text — released by Zhipu on 11 August 2025.^[2] It is a fine-tune of GLM-4.5-Air-Base, and so sits on top of GLM-4.5-Air, a text model with 106B total parameters, of which 12B are active.^[1] That distinction matters: the model as a whole holds a large amount of stored knowledge, while only a fraction of its parameters is used at any one step. The model card describes GLM-4.5-Air as ZhipuAI's next-generation flagship text foundation model.^[1]

Where it comes from

GLM-4.5V belongs to Zhipu's GLM-V line of multimodal models. The technical report covering it and its predecessor, arXiv 2507.01006, was submitted on 1 July 2025 by the GLM-V Team.^[4] In the line's timeline the model follows GLM-4.1V-9B-Thinking and comes before GLM-4.6V, released in December 2025.^[3][2] The report notes that both GLM-4.1V-9B-Thinking and GLM-4.5V were open-sourced.^[4]

What it can do

The model is built for a broad set of multimodal tasks: image reasoning, video understanding, GUI (graphical user interface) tasks, parsing of complex charts and long documents, and visual grounding — locating the part of an image that a piece of text refers to.^[1] Video understanding covers long-video segmentation and event recognition, and the GUI work includes screen reading, icon recognition and assistance with desktop operations.^[1]

Like GLM-4.5, it has a thinking-mode switch.^[1] The idea behind such a switch is that a model can either answer straight away or first write out a longer chain of reasoning and then answer; the switch lets the user decide which is worth the extra time for a given task.

How it was trained

The technical report says the model used reinforcement learning with curriculum sampling (RLCS) to improve its performance on STEM problem solving, video understanding, coding, grounding and GUI-based agents.^[4] In reinforcement learning the model generates candidate answers, receives a reward for how good they are, and is nudged toward the answers that score higher. Curriculum sampling refers to how training problems are chosen and ordered — a planned progression rather than one fixed mixture kept throughout training.

Results

Zhipu said the model reached state-of-the-art performance among models of its scale on 42 public vision-language benchmarks.^[1] The technical report says GLM-4.5V is competitive with, or superior to, Gemini-2.5-Flash on tasks such as coding and GUI agents.^[4] These are the developers' own comparisons, made at release time.

Where it sits

GLM-4.5V is released under the MIT License, which places few restrictions on use and modification — a practical point for teams that want to run a multimodal model on their own infrastructure.^[1] Within Zhipu's line it stands between GLM-4.1V-9B-Thinking and GLM-4.6V, one step in a sequence of open multimodal releases rather than a single isolated model.^[3][2] Its feature list — reasoning over images, long documents, video and interface control in one model — is a fair picture of what vision-language models were being asked to do in this period.