Zhipu releases the GLM-4.5V vision-language model

What was released

On 11 August 2025, Zhipu released GLM-4.5V, a vision-language model: a system that takes images or video together with text as input and answers in text.^[2] The release is an open one, and it comes with a published paper describing the model and its training approach.^[3]

Vision-language models extend a text-only language model with a second input channel. In general terms, such a model learns to connect visual features with words, so that a question about a picture can be answered in much the same way as a question about a paragraph. That connection is what lets a single model handle jobs that would otherwise need a pipeline of separate programs: reading a chart, describing a video, or pointing to the part of an image that a sentence refers to.

What it is built on

GLM-4.5V is built on GLM-4.5-Air and is a fine-tune of GLM-4.5-Air-Base.^[1] GLM-4.5-Air has 106 billion total parameters, of which 12 billion are active.^[1] The two figures measure different things. The total counts everything the model stores; the active count is what actually gets computed for a given piece of input. A large gap between them reflects a sparse design, in which only part of the network runs for each token, so the cost of a forward pass stays closer to the smaller number while the model keeps the capacity of the larger one.

A fine-tune, in turn, is a model produced by continuing training from an already trained base model rather than starting from scratch. Here the base is GLM-4.5-Air-Base, and GLM-4.5V is the result of adapting it further.^[1]

What it can do

The capabilities Zhipu lists for GLM-4.5V span several kinds of visual work: image reasoning, video understanding, GUI tasks, parsing of complex charts and long documents, and visual grounding.^[1] Some of these are worth unpacking. GUI tasks means working with a software interface as a person would — reading what is on screen and acting on it — which is a different skill from describing a photograph, because the model has to locate small elements such as buttons and fields and connect them to a goal. Visual grounding means tying a piece of text to the region of an image it refers to: given a phrase, the model has to say where in the picture the thing named by that phrase is. Chart and long-document parsing is the reading of dense, structured layouts where the information sits in tables, axes and small print rather than in ordinary prose.

Reported results

Zhipu said that GLM-4.5V reached state-of-the-art performance among models of its scale on 42 public vision-language benchmarks.^[1] The accompanying paper states the same result in a slightly different form: state-of-the-art performance on nearly all tasks among open-source models of similar size across 42 benchmarks.^[3] The paper also reports that GLM-4.5V is competitive with or superior to Gemini-2.5-Flash on coding and GUI-agent tasks.^[3] These are the developer's and the authors' own reported comparisons, made on a selected set of benchmarks; whether they hold up under independent testing is a separate question that outside evaluations would have to settle.

License and companion release

GLM-4.5V is released under the MIT License.^[1] That is a permissive open-source license: it allows use, modification and redistribution of the software, including in commercial products, provided the license notice is kept. For a model of this size, that terms choice matters as much as the benchmark numbers, because it decides who can legally download the weights, change them and ship them inside another product.

The paper notes that GLM-4.5V was open-sourced alongside GLM-4.1V-9B-Thinking, a second, much smaller vision-language model released at the same time.^[3]

Why it matters

The interest in GLM-4.5V comes from where two things meet: a vision-language model capable of GUI and coding work, and an open license. Models that can read an interface and act on it are the part of multimodal research closest to software agents, since an agent that operates programs has to understand screens rather than plain text.^[3] Publishing the weights under MIT puts that capability in the hands of researchers and smaller teams who cannot train such a model themselves, and it lets outside groups inspect and test the claims in the paper directly.^[1] The reported parity or advantage against Gemini-2.5-Flash on coding and GUI-agent tasks, if it survives scrutiny, would place an openly licensed model in the same conversation as a widely used proprietary one on exactly those tasks.^[3]