MiroFlow v0.2

MiroFlow v0.2: A High-Performance and Cost-Effective Open-Source Framework for Research Agents

Research Agent Demo is live - Try Now

TL;DR

MiroFlow is a powerful open-source research agent framework that can upgrade any LLM to OpenAI Deep Research-level capabilities. It is a key component of the MiroMind Research Agent Project, and is designed to reliably complete complex tool-use tasks.

Key Features

• Reproducible State-of-the-Art Performance: Open-source, reproducible SoTA results across GAIA, HLE, xBench-DeepSearch, and BrowserComp benchmarks.

• High Concurrency & Reliability: Built with robust concurrency management and fault-tolerant design, MiroFlow efficiently handles rate-limited APIs and unstable networks, ensuring seamless trajectory collection and reliable execution of complex tasks.

• Cost-Effective Deployment: Powered by the open-source MiroThinker model, MiroFlow can run a research agent service on a single RTX 4090. The entire stack relies on free, open-source tools, making it simple to deploy, scale, and reproduce.

1. Overall Architecture

• Frontend

A simple Gradio frontend

• Backend

MiroFlow automatically handles user queries through multi-tool collaboration (such as web browsers and Python tools), performing multi-step web research, comprehensively analyzing a large number of online resources, and ultimately completing the task. The specific process includes:

Query Augmentation: The user input is first analyzed by a large language model (LLM) to identify the user’s intent and enrich query details, enabling a more accurate understanding of the requirements.

Task Planning: The main agent formulates a detailed execution plan based on the enhanced query content, coordinating the entire workflow, including invoking different tools, assigning tasks to sub-agents, and driving task progress.

Sub-Agent Delegation: For complex or specialized tasks, the main agent delegates parts of the work to sub-agents with relevant expertise (e.g., agent-browsing). These sub-agents can independently plan and execute tasks, as well as call upon necessary tools.

Tool Calling: When external functionalities need to be invoked, agents connect to the MCP (Model Context Protocol) server to obtain and use the corresponding specialized tools.

Result Synthesis: After task completion, the system consolidates results from multiple information sources to ensure the output is of high quality and meets user requirements or preset formats.

2. MiroFlow v0.2 vs. Existing Research Agent Frameworks (2025-08-21)

• One Framework for Multiple Benchmarks: Supports GAIA, HLE, BrowserComp, and xBench-DeepSearch.

• Reproducible State-of-the-Art Performance: MiroFlow delivers open-source state-of-the-art results across the representative benchmarks, including GAIA, xBench, BrowserComp, and HLE . Unlike commercial frameworks or partially open-sourced research efforts, every reported metric is fully reproducible with the publicly available code: github.com/MiroMindAI/MiroFlow.

“Open-Source Reproducible” means that the model, code, and runtime environment are all open source, and by running the provided test scripts, the experimental metric results can be reproduced.

“*” means: the model is evaluated on the GAIA text-103 subset.

Gemini Deep Research “50.0+” on xBench-DeepSearch is from the Kimi Researcher report.

• Cost-Effective Deployment: Powered by the open-source MiroThinker model, MiroFlow can run a research agent service on a single RTX 4090. The entire stack relies on free, open-source tools, making it simple to deploy, scale, and reproduce. Code available: https://github.com/MiroMindAI/MiroThinker.