OpenViSTAI · Agents · Blockchain · Crypto
Submit a project
Menu
AI Models

Glm 4.5

GLM-4.5 is a family of foundation language models designed for intelligent agents, combining reasoning, coding and agent capabilities. GLM-4.5 and GLM-4.5-Air offer thinking and non-thinking modes, and the repository identifies model downloads and Z.ai API access.

Explore the practical guide ↓

Independent OpenViST project profile, based on linked sources.

Glm 4.5 — OpenViST project profile in AI Models.
OpenViST project illustration. See the linked sources for product details.
Understand the project

Practical overview

When it fits

Consider GLM-4.5 when building an application that needs a reasoning and coding model for agent tasks, and choose between downloaded model artifacts and the API route identified by the repository.

Who it is for

  • Developers building coding assistants or reasoning-based question-answering applications.
  • Teams prototyping agents that need reasoning, coding or tool-use capabilities.

Documented uses

  • Add a language model to a coding assistant.
  • Power question answering that relies on reasoning.
  • Prototype an agentic task assistant.
Before you choose

Requirements & limits

The repository identifies API services through the Z.ai API Platform and lists open model releases. It describes the GLM-4.5 model releases as MIT-licensed; repository metadata lists Apache-2.0, so check the license for the specific artifact being used.

  • GLM-4.5 has 355 billion total parameters and 32 billion active; GLM-4.5-Air has 106 billion total and 12 billion active.
  • Both variants are described as supporting thinking mode for complex reasoning and tool use, and non-thinking mode for immediate responses.
  • The repository points to both downloadable model artifacts and API access.
  • The README describes MIT licensing for model releases while repository metadata lists Apache-2.0; confirm the applicable license for the selected artifact.

Check current requirements in the official documentation. Source-based guidance does not mean OpenViST has run or independently tested this project.

Source-based practical guide

Start GLM-4.7 with vLLM

Serve the GLM-4.7-FP8 model using the documented vLLM configuration.

Before you start

  • Install the dependencies, including vLLM, according to the repository’s requirements.txt configuration.
  • Use a setup that meets the repository’s inference requirements: FP8-capable devices, server memory exceeding 1 TB, and an inference batch size no greater than 8. The listed GLM-4.7-Flash GPU configurations do not establish requirements for GLM-4.7.

Steps

  1. Install dependencies according to requirements.txt.

  2. Run the repository’s documented vLLM serve command for zai-org/GLM-4.7-FP8, with tensor parallel size 4, the MTP speculative configuration, glm47 tool-call parser, glm45 reasoning parser, automatic tool choice enabled, and served model name glm-4.7-fp8.

Expected outcome

The documented vLLM command starts a server configured to serve GLM-4.7-FP8 under the name glm-4.7-fp8.

Guide limitations

  • The README provides this command for GLM-4.7-FP8, not GLM-4.5.
  • The documented GPU and memory tables list GLM-4.5, GLM-4.5-Air, and GLM-4.7-Flash configurations, not GLM-4.7-FP8; they do not confirm that this command will run on a particular setup.
  • The README does not provide a request example or describe a successful test.
Documentation 1 ↗
Source-based practical guide

Launch GLM-4.5-Air with SGLang PD disaggregation

Start separate prefill and decode servers for GLM-4.5-Air and connect them with the documented SGLang router command.

Before you start

  • Install the dependencies, including SGLang, according to the repository’s requirements.txt configuration.
  • Use one machine with multiple GPUs; the README describes P and D using four GPUs each for this method.
  • Use the documented interconnect device mlx5_0.

Steps

  1. Start the prefill server with the documented SGLang command for zai-org/GLM-4.5-Air, using disaggregation mode prefill and tensor parallel size 4.

  2. Start the decode server with the documented command, using disaggregation mode decode, port 30001, tensor parallel size 4, and base GPU ID 4.

  3. Start the SGLang router with the documented PD-disaggregation command, pointing prefill to http://127.0.0.1:30000 and decode to http://127.0.0.1:30001, and setting its port to 8000.

Expected outcome

The documented commands launch prefill and decode processes and a router configured to connect them.

Guide limitations

  • The README describes this as a simple single-machine method and specifies four GPUs each for P and D.
  • No request example or successful test is provided.
  • The general inference requirements specify server memory exceeding 1 TB, batch size no greater than 8, and FP8-capable devices, but the README does not clarify how these conditions apply to this PD-disaggregation configuration.
Documentation 1 ↗
Evidence and transparency

Sources & review

Official repository ↗
Source record details
Source availability
Active
Sources checked
7 October 2026
Original website title
zai-org/GLM-4.5 · Official homepage title · retained in its source language
How we review sourcesSuggest a correction