Supported Models & Setup

Relevant source files

The following files were used as context for generating this wiki page:

• src/utils/llm.py
• README.md
• src/utils/utils.py
• tests/test_llm_api.py
• webui.py
• src/utils/default_config_settings.py

The Supported Models & Setup module defines the integration layer between the Browser Use Web UI and various Large Language Model (LLM) providers. This system enables the AI agent to utilize a diverse range of models, including proprietary APIs and local instances, to process user instructions, analyze browser states, and determine next actions.

The architecture supports multiple providers such as OpenAI, Anthropic, Google, and DeepSeek, alongside local hosting via Ollama. It manages configuration via environment variables or direct UI input, handling model-specific logic like reasoning content extraction for DeepSeek-R1 and context window management for local models.

Sources: README.md:15-18, src/utils/utils.py:42-140

LLM Provider Architecture

The project employs a factory-style pattern to initialize LLM clients using the get_llm_model function. This function abstracts the complexities of different provider SDKs (OpenAI, Anthropic, Google, Mistral, etc.) and returns a unified LangChain compatible chat model object.

The diagram shows the flow from a request to the get_llm_model factory which routes to specific provider implementations.
Sources: src/utils/utils.py:42-138, tests/test_llm_api.py:35-51

Specialized Model Implementations

For models that require custom handling of response metadata (like reasoning chains), the system implements specialized classes. Specifically, DeepSeekR1ChatOpenAI and DeepSeekR1ChatOllama are used to extract "reasoning_content" or content within <think> tags.

• Reasoning Extraction: In DeepSeekR1ChatOllama, the invoke method splits the raw content by the </think> tag to separate the agent's internal reasoning from its final output.
• Context Management: For Ollama models, a num_ctx parameter is utilized to control the maximum context length, which defaults to 32,000 but can be adjusted via the UI.

Sources: src/utils/llm.py:53-128, src/utils/utils.py:108-124

Supported Providers & Models

The following table lists the primary supported providers and their default or common models as defined in the source code.

Provider	Default/Key Models	Configuration Key	Source
OpenAI	gpt-4o, gpt-4, gpt-3.5-turbo, o3-mini	OPENAI_API_KEY	utils.py:144
Anthropic	claude-3-5-sonnet, claude-3-opus	ANTHROPIC_API_KEY	utils.py:143
Google	gemini-2.0-flash, gemini-1.5-flash	GOOGLE_API_KEY	utils.py:146
DeepSeek	deepseek-chat, deepseek-reasoner	DEEPSEEK_API_KEY	utils.py:145
Ollama	qwen2.5, llama2, deepseek-r1	N/A (Local)	utils.py:147
Azure OpenAI	gpt-4o, gpt-4	AZURE_OPENAI_API_KEY	utils.py:148
Mistral	pixtral-large-latest, mistral-large	MISTRAL_API_KEY	utils.py:149

Sources: src/utils/utils.py:142-152, tests/test_llm_api.py:54-62

Setup & Configuration

Configuration is managed through a combination of environment variables and a persistent settings system.

Environment Configuration

Users can define global defaults in a .env file. The application maps these to specific provider settings during initialization.

env_mappings = {
    "openai": {"api_key": "OPENAI_API_KEY", "base_url": "OPENAI_ENDPOINT"},
    "azure_openai": {"api_key": "AZURE_OPENAI_API_KEY", "base_url": "AZURE_OPENAI_ENDPOINT"},
    "google": {"api_key": "GOOGLE_API_KEY"},
    "deepseek": {"api_key": "DEEPSEEK_API_KEY", "base_url": "DEEPSEEK_ENDPOINT"}
}

Sources: tests/test_llm_api.py:54-62, README.md:144-150

Web UI Settings

The Web UI provides a configuration tab where users can load and save settings as .pkl files. The default_config function establishes the baseline parameters for the agent and LLM.

• Temperature: Controls randomness (default: 1.0).
• Max Context Length: Specific to local models like Ollama.
• Base URL: Allows redirection to custom endpoints or local proxies (e.g., http://localhost:11434 for Ollama).

Sources: src/utils/default_config_settings.py:8-32, webui.py:465-490

Initialization Logic

The initialization of the LLM involves resolving the API key, setting the base URL, and instantiating the correct LangChain class.

The sequence diagram illustrates how the get_llm_model function validates inputs and selects the specialized or standard LangChain class.
Sources: src/utils/utils.py:42-140, src/utils/llm.py:53-128

Summary

The "Supported Models & Setup" system provides a robust abstraction layer for LLM integration. By combining standard LangChain classes with custom extensions for DeepSeek reasoning and Ollama context management, it ensures that the browser agent can operate across a wide spectrum of AI backends. Configuration is flexible, allowing for quick deployment via environment variables or granular control through the Gradio-based Web UI.