GLM-5V-Turbo

GLM-5V-Turbo, released at the start of April 2026, is the model with which Z.ai joined vision to programming. Where the text models of the GLM-5 family read specifications, this one reads pictures: design mockups, wireframes, screenshots of broken pages, video, whole documents. The output is always text — in practice, code. The use cases the company documents say more about the model than any benchmark table. From a mockup it reconstructs layout, palette, component hierarchy and interaction logic, and returns a runnable front-end project. From a screenshot of a misrendered page it identifies overlapping components, misaligned layout and wrong colours, and proposes a fix. Working inside an agent it goes further and browses the target site on its own, maps transitions between pages, collects visual assets — Z.ai calls this the step from "recreating from a screenshot" to "recreating through autonomous exploration". The same perception is what lets it drive graphical interfaces, which the company measures on AndroidWorld and WebVoyager. Technically it takes video, images, text and files in, holds 200,000 tokens of context and returns up to 128,000, with thinking modes, streaming, function calling and context caching. Z.ai describes a CogViT vision encoder, an inference-friendly multi-token-prediction architecture and reinforcement learning run jointly across more than thirty task types. The price is $1.20 per million input tokens and $4.00 per million output — the same tier as the text model GLM-5-Turbo. Two absences are worth naming. Z.ai states no parameter count, and unlike the text line it did not publish the weights: GLM-5V-Turbo is available through the API only, so the usual check against the safetensors index is not possible here. The company repeatedly stresses that the model achieves its results "with a smaller model size" — a claim that, without a published number, cannot be verified.

▸Videos

No videos yet.

▸Reviews