glm5.2 does not show the thought process

[leibHz]

Description

I am using GLM 5.2 with NVIDIA as the provider via NVIDIA NIM, whereas other models indicate they are thinking by displaying a "thinking" notification, The GLM simply isn't showing anything; I suspect the parameters are being sent incorrectly, given the 'default', 'high', and 'max' variants—variants that NVIDIA NIM doesn't actually support. It should just be enable_thinking=true, at least based on what I saw in their documentation.

He's also responding very quickly, which leads me to believe he isn't even thinking.

Plugins

gsd

OpenCode version

1.18.9

Steps to reproduce

1. select glm 5.2 with the nvidia provider

2. ask/request anything

Screenshot and/or share link

Operating System

Linux 7.1.5-1-cachyos

Terminal

fish 4.8.1 on vscode

[github-actions]

This issue might be a duplicate of existing issues. Please check:

• Minmax and GLM models (Minmax 2.7, GLM 5,5.1 ) no thinking level support via opencode zen provider #24010: GLM models (GLM 5, 5.1) also show no thinking level support — same family of models with the same symptom
• Nvidia NIM API hangs for DeepSeek v4 reasoning models without chat_template_kwargs #24264: NVIDIA NIM requires chat_template_kwargs: { enable_thinking: true } for reasoning models — directly related to the incorrect thinking parameters you describe for NVIDIA NIM

If your issue is specifically about the wrong thinking parameters being sent to NVIDIA NIM for GLM 5.2 (e.g., sending default/high/max variants instead of enable_thinking=true), it may overlap with #24264 which covers the same NVIDIA NIM chat_template_kwargs requirement.

[lenmei233]

I've submitted a PR (#39569) to fix this properly in the codebase. While waiting for it to be reviewed and merged, you can use the following workaround by configuring a custom OpenAI-compatible provider in your OpenCode config file (~/.config/opencode/opencode.json or project-level .opencode.json):

{
  "$schema": "https://opencode.ai/config.json",
  "provider": {
    "nvidia-glm": {
      "npm": "@ai-sdk/openai-compatible",
      "name": "NVIDIA NIM (GLM)",
      "options": {
        "baseURL": "https://integrate.api.nvidia.com/v1",
        "apiKey": "{env:NVIDIA_API_KEY}"
      },
      "models": {
        "z-ai/glm-5.2": {
          "name": "GLM 5.2 (thinking)",
          "interleaved": {
            "field": "reasoning_content"
          },
          "options": {
            "chat_template_kwargs": {
              "enable_thinking": true,
              "clear_thinking": false
            }
          }
        }
      }
    }
  }
}

Then select nvidia-glm/z-ai/glm-5.2 from the model picker in OpenCode. Make sure to set the NVIDIA_API_KEY environment variable.

The key points:

• Uses @ai-sdk/openai-compatible instead of the built-in nvidia provider to avoid sending the unsupported reasoningEffort parameter.
• Injects chat_template_kwargs directly via options, which is what NVIDIA NIM expects for GLM models.
• Sets interleaved.field to reasoning_content so the thinking stream appears in the UI.

You can also verify the fix works by testing the raw API call with curl:

curl https://integrate.api.nvidia.com/v1/chat/completions \
  -H "Authorization: Bearer $NVIDIA_API_KEY" \
  -H "Content-Type: application/json" \
  -d '{
    "model": "z-ai/glm-5.2",
    "messages": [{"role": "user", "content": "Prove sqrt(2) is irrational"}],
    "stream": true,
    "chat_template_kwargs": {"enable_thinking": true, "clear_thinking": false}
  }'

If you see reasoning_content in the stream, the workaround works correctly.

My PR (#39569) implements the same logic inside the built-in NVIDIA provider, so once merged, no manual configuration will be needed.