MiroThinker
Overview
MiroThinker is the official open-source implementation of the MiroMind research agent project. It focuses on building search/research agents that natively support tool-augmented reasoning and deep multi-step interactions with external environments (web search, code execution, scraping, vision/audio tools, etc.). The project bundles model releases, an agent framework (MiroFlow), a curated training dataset (MiroVerse), and training/serving infrastructure (MiroTrain / MiroRL) to enable reproducible research and real-world research workflows.
Key components
- MiroThinker (agent models): model series released at multiple parameter scales (examples include v1.0 and v1.5 families). Later releases (v1.5) emphasize "interactive scaling," very long context windows (256K), and high tool-call budgets to support deep, long-horizon reasoning.
- MiroFlow: the open-source agent framework that orchestrates tools, logging, context management, and benchmark evaluation to reproduce state-of-the-art results.
- MiroVerse: an open training dataset (about 147k samples in v0.1) tailored to train research agents with realistic tool-usage traces and interactions.
- MiroTrain / MiroRL: training and reinforcement-learning infrastructure for stable, efficient training of agent models.
Features & Capabilities
- Tool-augmented reasoning: native integration with web search, scraping, code execution sandboxes, VQA/transcription tools and customizable MCP servers.
- Interactive scaling: training the model to handle deeper and higher-frequency agent-environment interactions as an orthogonal dimension to model size and context length.
- Long context support: models and workflows supporting up to 256K context windows (notably in v1.5).
- High tool-call budgets: variants configured to handle hundreds of tool calls per task (v1.0 reported up to 600 in some configs; v1.5 commonly up to 400 in the repo documentation).
- Reproducible benchmark suite: extensive evaluation scripts and configs for many benchmarks (GAIA, HLE, BrowseComp, XBench-DeepSearch, FutureX, Frames, SEAL-0, AIME2025, etc.).
Models & Performance (as reported in the repository)
The README documents multiple model releases and reported benchmark numbers. For example, MiroThinker-v1.5 is released in 30B and 235B variants with 256K context and high tool-call budgets, and the repo reports strong results across BrowseComp, HLE-Text, and GAIA validation tasks. Earlier series (v1.0, v0.2, v0.1) are retained for reference and comparisons.
Quick start & deployment
The project provides step-by-step instructions for cloning, environment setup, configuring required API keys (Serper, Jina, E2B, optional OpenAI/Anthropic for evaluation), and running the agent using provided configurations. It also documents recommended agent configurations for different benchmarks (context-management presets) and several serving options (SGLang, vLLM, quantized CPU/GPU options).
Tools & Integration
MiroThinker ships with a configurable set of MCP servers and tool adapters (search, scraping, sandboxed code execution, VQA/transcribe, reasoning modules). The framework encourages mixing open-source and commercial tools for research and benchmarking while documenting how to reproduce reported results with open configurations.
Community, license & reproducibility
The repository is MIT-licensed and links to model and data artifacts on Hugging Face (models, MiroVerse dataset). The project provides trace collection scripts, benchmark evaluation scripts, monitoring utilities, and community links (Discord, blog, issues) to facilitate reproducible research and community contributions.
Who should use it
Researchers and engineers building or evaluating tool-augmented research agents, teams that need reproducible agent benchmarks, and practitioners who want a full-stack open-source agent framework (models, data, training, and tools) for long-context, multi-step research tasks.
