GLM-4.5-Air Locally on WSL With Ollama

GLM-4.5-Air is an older open-weight GLM model that can run locally, but “Air” does not mean small. Z.AI lists 106 billion total parameters with 12 billion active parameters. A popular community Ollama build ranges from about 40 GB to 117 GB, depending on quantization. Check your available memory before downloading it.

This covers the local run on WSL and connecting it to an agent. For the far more capable hosted option, see GLM Coding Plan setup.

Why GLM-4.5-Air for local#

Step 1: Ollama on WSL#

If WSL isn’t installed, start with the WSL install guide. Install Ollama inside WSL (or use native Windows Ollama). It serves models on a local OpenAI-compatible endpoint at http://localhost:11434/v1.

Step 2: Get GLM-4.5-Air into Ollama#

Check Ollama’s model library for a GLM-4.5-Air build. If one is listed:

ollama pull MichelRosselli/GLM-4.5-Air:IQ1_M

If it isn’t in the library, you can import a compatible GGUF build via a Modelfile per Ollama’s docs. Confirm the model name you end up with, as you’ll need it for the agent config.

Step 3: Test it#

ollama run MichelRosselli/GLM-4.5-Air:IQ1_M

Give it a small coding question to confirm it loads and responds at a usable speed.

Step 4: Connect a coding agent#

Claude Code via Claude Code Router:

{
  "Providers": [
    {
      "name": "ollama",
      "api_base_url": "http://localhost:11434/v1/chat/completions",
      "api_key": "ollama",
      "models": ["MichelRosselli/GLM-4.5-Air:IQ1_M"]
    }
  ],
  "Router": { "default": "ollama,MichelRosselli/GLM-4.5-Air:IQ1_M" }
}

Then ccr code.

Cline in VS Code can use http://localhost:11434/v1 through its OpenAI Compatible provider. Enter the exact Ollama model name you downloaded.

The honest trade-offs#

Air is a lightweight model, so local quality is well below the hosted GLM-5/Turbo on the plan. This setup is about privacy and zero cost, not capability. For real coding power at low cost, the GLM Coding Plan is hard to beat at $18/month.

Troubleshooting#

• If Ollama reports that the model is missing, use the full community namespace and tag.
• If loading fails, choose a smaller quantization or use a machine with more memory.
• If the agent cannot connect, confirm that Ollama is running on port 11434.
• If tool calls fail, remember that the community conversion’s template may differ from Z.AI’s reference runtime.

Local or hosted#

Local GLM-4.5-Air is useful when data must remain on your machine and you have enough memory for a large quantized model. For a new hosted coding setup, use GLM-5.3 instead.

For other local routes, see DeepSeek locally with Ollama and Qwen3-Coder locally with LM Studio.