<MCP.Playground/>
online

GLM 5V Turbo

Runs MCP tools

Z.ai’s first native multimodal agent model, released April 2026. It is built to look at a screen or a design, plan, and act on it, which makes it a fit for MCP servers that return screenshots.

Vendor
Z.ai
Released
1 April 2026
Context
203K tokens
Input
Text, image, video

Paste a server URL, pick a model, and watch it call your tools in a real conversation. No install.

No server of your own? Leave it blank and use one of the public mock servers.

MCP Evals

Can GLM 5V Turbo actually complete tasks with your server?

One chat shows the tool calls work. An eval writes real tasks from your tool schemas, has GLM 5V Turbo drive each one, and reports where it got the wrong answer even though every call succeeded. Pick GLM 5V Turbo as the driver.

Run evals →See a sample

What GLM 5V Turbo is

GLM 5V Turbo pairs GLM-5-Turbo with CogViT, a vision encoder Z.ai built for it. It takes text, image and video input, has a 200K window and up to 128K tokens of output, and can think before it answers. Z.ai built it for vision-based coding and GUI agents: turning designs into front-end code, operating Android and desktop interfaces, and debugging from screenshots. Z.ai names Claude Code and OpenClaw as agents it is meant to work in.

What Z.ai says

  • Design2Code: 94.8.
  • AndroidWorld: 75.7. OSWorld: 62.3.
  • BrowseComp-VL: 51.9.
  • CC-Bench-V2: 68.4 on front-end tasks, and 22.8 on back-end tasks.

These are the vendor’s own claims, not measurements of ours. Run the model against your own server to find out whether they hold for your tools.

How it reaches MCP

  • Claude Code and OpenClaw, which Z.ai names as supported agents. Both are MCP clients.
  • Z.ai’s OpenAI-compatible API as glm-5v-turbo, so MCP clients that accept a custom endpoint can point at it with a base-URL change
  • MCP Playground’s Agent Studio: paste your server URL above and run GLM 5V Turbo against it in the browser, with no Z.ai key

What to watch for

  • Its strength is what it sees. It pays off with MCP servers that return image content, such as browser and screenshot tools, and only if your client passes those images to the model rather than dropping them.
  • The back-end coding score is weak: 22.8 on CC-Bench-V2, against 68.4 on front-end. For text-only servers, GLM 5.3 or GLM 5.3 Flash is the better GLM.
  • Z.ai’s own paper names multimodal context management as the main bottleneck for long-horizon agents. Screenshots fill the 200K window fast, so a long GUI session is where to look for it losing track.

Run GLM 5V Turbo and GLM 5.3 Flash on the complex-schema mock with the same prompt. Both take images, but 5V Turbo was built for them. If your server returns images, give it a task that needs one, and check whether it uses what the image shows. All the mock servers →

Available here

ModelModel ID
GLM 5V Turboz-ai/glm-5v-turbo

Go deeper

Frequently asked questions

What is GLM 5V Turbo?
Z.ai’s first native multimodal agent model, built on GLM-5-Turbo with a new vision encoder. It is aimed at design-to-code work and agents that operate a screen.
Does GLM 5V Turbo support tool calling?
Yes. OpenRouter lists tools and tool_choice, and Z.ai designed it for plan-and-act agent loops. It does not list parallel_tool_calls, structured_outputs or seed.
GLM 5V Turbo or GLM 5.3 Flash?
Both take images. 5V Turbo was built for vision-led agent work and scores highly on GUI and design-to-code benchmarks. 5.3 Flash is newer, has a window of over 1M tokens, and is the better general choice for text-only servers.
Can I use GLM 5V Turbo with Claude Code?
Z.ai lists Claude Code as a supported agent, and Claude Code is an MCP client, so your MCP servers come with it.

Sources

GLM 5V Turbo for MCP: Specs & Tool Calling — Test Free | MCP Playground