Abstract:

This document announces the launch of Ollama's new web search API, designed to augment large language models with real-time information from the internet. The primary goal of this feature is to enhance model accuracy and reduce hallucinations by providing up-to-date context. The capability is delivered as a REST API and is deeply integrated into Ollama's Python and JavaScript libraries, enabling functions like web_search for general queries and web_fetch for retrieving content from specific URLs.

The announcement showcases how to use the API as a tool to build a search agent, which can conduct multi-step research. An example using the qwen3 model reveals other significant 2025 engine updates, including an improved model scheduling system for better performance and memory management, and a new multimodal engine with support for vision models. The API is available to all users with an Ollama account, offering a free tier for individuals and higher rate limits for subscribers.

Ollama Unveils New Web Search API and Engine Enhancements

• New Web Search API: Ollama has released a new web search capability to connect its models with live internet data. This feature aims to improve accuracy and reduce model hallucinations.
• Simple API Access: The functionality is provided as a REST API that can be accessed via cURL or integrated libraries. Users need to generate an OLLAMA_API_KEY from their account to get started.
• Python and JavaScript Integration: The API is deeply integrated into Ollama's official libraries (ollama>=0.6.0). Users can easily call the search function using ollama.web_search() in Python or client.webSearch() in JavaScript.
• Building Search Agents: The web search can be used as a tool within models to create powerful search agents. An example is provided showing a qwen3:4b model using the tool to answer a user's query about Ollama's new engine.
• Enhanced Model Scheduling Engine: The search agent example reveals a major engine update (September 23, 2025) that includes precision memory management to reduce crashes, improved GPU utilization, and better multi-GPU support, leading to significant performance gains.
• New Multimodal Engine: Another key update is a multimodal engine (May 15, 2025) that adds first-class support for vision models like llama4:scout and gemma3, enabling tasks such as image and video analysis.
• Dedicated Web Page Fetching: In addition to broad searches, a new web_fetch API allows models to retrieve and process the content of a specific URL provided by a user.
• Recommended Models: For best performance with tool use, Ollama recommends models like qwen3 and gpt-oss. It is also suggested to use models with a context length of at least 32,000 tokens.
• Third-Party Integrations: The web search functionality can be integrated into various developer tools, including MCP Server, Cline, Codex, and Goose.
• Availability: The web search API is available to all users with a free Ollama account. Higher rate limits and advanced features are available through a paid subscription.