Composed Models

Some models excel at grounding (finding UI elements) but can't reason about tasks. Others are great at planning but can't precisely locate elements. Composed models combine both capabilities for better results.

How It Works

Use the + syntax to combine a grounding model with a planning model:

from computer import Computer
from agent import ComputerAgent

computer = Computer(os_type="linux", provider_type="docker", image="trycua/cua-xfce:latest")
await computer.run()

# Format: grounding_model+planning_model
agent = ComputerAgent(
    model="huggingface-local/HelloKKMe/GTA1-7B+anthropic/claude-sonnet-4-5-20250929",
    tools=[computer]
)

Cua uses a two-stage approach:

1. Grounding stage - The first model analyzes the screenshot and identifies UI element coordinates
2. Planning stage - The second model receives the element list and decides which to interact with

Common Combinations

Local Grounding + Cloud Planning

Run visual processing locally while using cloud models for reasoning:

# GTA1 for grounding + Claude for planning
agent = ComputerAgent(
    model="huggingface-local/HelloKKMe/GTA1-7B+anthropic/claude-sonnet-4-5-20250929",
    tools=[computer]
)

# Moondream3 for grounding + Claude for planning
agent = ComputerAgent(
    model="moondream3+anthropic/claude-sonnet-4-5-20250929",
    tools=[computer]
)

# OmniParser for grounding + GPT-4o for planning
agent = ComputerAgent(
    model="omniparser+openai/gpt-4o",
    tools=[computer]
)

# UI-TARS for grounding + GPT-4o for planning
agent = ComputerAgent(
    model="huggingface-local/ByteDance-Seed/UI-TARS-1.5-7B+openai/gpt-4o",
    tools=[computer]
)

Grounding Models

These models locate UI elements but can't reason about actions:

Planning Models

Any full computer-use model can serve as a planner:

When to Use Composed Models

Privacy - Run grounding locally so screenshots never leave your machine. Only text descriptions of UI elements go to the cloud.

Cost - Grounding models process images locally, reducing API calls to cloud providers.

Latency - Local grounding can be faster than sending full screenshots over the network.

Accuracy - Some grounding models detect UI elements more precisely than general-purpose VLMs.

Limitations

• Requires a local GPU for the grounding model (typically 16GB+ VRAM for 7B models)
• Two-stage processing adds some overhead
• Not all grounding models work with all planning models

For simpler setups, consider using a single full computer-use model like Claude or OpenAI computer-use-preview.