Run AutoGLM-Phone-9B-Multilingual for English apps

If your target apps are in English, the AutoGLM‑Phone‑9B‑Multilingual model is typically the safer default. This tutorial explains when to choose it, how to download it from official sources, and how to validate English UI behavior.

Why the multilingual model for English?

Even for English‑only apps, multilingual models often have broader UI text coverage and more robust handling of mixed‑language labels, localization menus, and embedded foreign strings. In practice, that reduces the chance of mis‑reading buttons or mis‑interpreting settings screens.

This is not a claim of superior performance for every use case. It is a practical default for evaluation.

Official sources only

Download the model from:

• Hugging Face: TODO add the official AutoGLM‑Phone‑9B‑Multilingual URL.
• ModelScope: TODO add the official AutoGLM‑Phone‑9B‑Multilingual URL.

Avoid re‑hosted weights. Official sources ensure consistent licensing and checksums.

Installation approach options

You can run the model in two main ways:

Option A (recommended): Use a deployed model service endpoint and pass --base-url.
Option B (local): Use a local GPU (24GB+ VRAM recommended) with vLLM or SGLang.

Option A is the fastest way to test whether the multilingual model behaves well on your UI flows.

Example configuration checklist

Before you run the agent:

• Confirm the model files and version.
• Store the model path and checksum in your test notes.
• Document device OS version, screen resolution, and target app versions.

These notes help you reproduce results later.

Example run sequence (template)

The official README is the source of truth. The sequence below is a template to show the structure, not exact commands.

# TODO: replace with official run command for the multilingual model
python -m openautoglm.run \
  --model /path/to/AutoGLM-Phone-9B-Multilingual \
  --device adb

If you are using an endpoint:

# TODO: replace with official endpoint flags
python -m openautoglm.run \
  --base-url https://your-endpoint.example.com

Prompt templates (safe defaults)

Use concise, deterministic prompts:

• "Open Settings and stop."
• "Scroll down until you see Accessibility, then stop and describe the screen."
• "Ask for confirmation before tapping any Save button."

These reduce ambiguous actions and keep the agent in safe territory.

Validating English UI flows

Start with low‑risk tasks:

1. Navigate to an app’s settings page.
2. Open the help or about screen.
3. Scroll a list and read visible labels.

These tasks help you evaluate perception accuracy and basic navigation without risk.

Prompting tips for English UI

To reduce ambiguity:

• Reference visible labels verbatim: “Tap Settings.”
• Use short, single‑action instructions.
• Ask the agent to describe the current screen before acting.

These small prompt changes can reduce mis‑clicks.

Handling mixed‑language screens

Even English apps can contain mixed UI strings (language selectors, locale switches, legal links). If the agent hesitates:

• Ask it to list all visible labels.
• Provide a screenshot and choose the target explicitly.
• Require a confirmation before final actions.

Common errors and how to handle them

• Mis‑read labels: add explicit text hints in prompts and confirm with screenshots.
• Unclear buttons: require the agent to ask for human confirmation before tapping.
• Language switching: ensure the device language is set to English consistently.

Safety and evaluation notes

Do not allow automated purchase or password‑change flows in early tests. Always set up a manual confirmation step for destructive actions.

Validation checklist

• Device language set to English
• Model version recorded
• Low‑risk tasks completed
• Confirmation prompts for risky actions

Lightweight evaluation rubric

Score each task:

• Accuracy: correct element identified (yes/no)
• Safety: no risky action taken (yes/no)
• Clarity: agent explanation was understandable (yes/no)

This keeps evaluations consistent across runs.

Example use cases (not endorsements)

• English onboarding flow validation.
• Regression checks on settings toggles.
• Copy and UI text verification before release.

Next steps

• Review the model download page for official sources.
• Compare with the low‑VRAM guide if you cannot run locally.