#Model Selection Guide

This guide introduces the models supported by Midscene, two ways to connect them, and when to use one Default model or combine multiple models.

#Supported models

Midscene supports these multimodal models for operating user interfaces: Doubao Seed, Qwen, DeepSeek, Google Gemini, OpenAI GPT, Moonshot Kimi, Xiaomi MiMo, Zhipu GLM-V, and Jev (experimental), as well as the other compatible models AutoGLM and UI-TARS.

#Choose a connection method

Midscene supports two ways to connect a model. Choose either one:

• Model API key: Use an API key from a model provider to connect to the models listed above. See Use a model API key.
• ChatGPT subscription: Sign in and authorize Midscene to call GPT models using your ChatGPT subscription. Available models depend on your account. See Use a ChatGPT subscription.

#Start with one Default model

Configure one multimodal Default model for planning, localization, and insight. This works for most scenarios. See Use a model API key for supported models and configuration examples, or choose Use a ChatGPT subscription.

#Use multi-model capabilities to handle complex scenarios

Midscene supports combining multiple models to handle complex scenarios. A single automation task may require multi-step planning, precise element localization, and data extraction or result evaluation. Different models have different strengths across these capabilities.

Keep the Default model and add Planning, Locate, or Insight models to divide the work according to those strengths within the same task. For example, use a stronger planning model for long workflows with conditional branches, a localization model for difficult UI targets, or an insight model for complex assertions and structured data extraction.

Configure each role independently; you do not need to add every specialized model at once. Any role without a dedicated model continues to use the Default model:

Combining models can increase latency and token usage. Use Model debugging and observability to identify the problem, then configure multiple models as needed.