MiroThinker is a deep research agent framework designed to handle complex research and prediction tasks through agentic reasoning and tool interaction. The repository, maintained primarily by contributors shawnlimn, Mereithhh, and Vanint according to GitGenius tracking data, represents an active project with a median issue and pull request response time of 43.3 hours across 66 tracked items. The codebase is written in Python and classified by GitGenius as spanning multiple AI domains including autonomous systems, cognitive AI, reasoning, planning, memory, and LLM integration.

The project has evolved through multiple major releases, with MiroThinker-1.7 representing the latest iteration as of March 2026. This version achieves 74.0 percent on BrowseComp and 75.3 percent on BrowseComp-ZH benchmarks, with the smaller 30-billion parameter variant MiroThinker-1.7-mini reaching 72.3 percent on BrowseComp-ZH while using significantly fewer parameters than competing models. The framework supports a 256K context window and can handle up to 300 tool interactions per task, enabling long-horizon reasoning and deep multi-step analysis across research workflows.

MiroThinker-1.7 comes in two parameter scales: a 30-billion parameter mini version and a 235-billion parameter full version, both available on Hugging Face. The larger model achieves 82.7 percent on GAIA-Val-165 and 42.9 percent on HLE-Text benchmarks. The repository also maintains documentation for previous versions, including MiroThinker-v1.5, which introduced interactive scaling as a performance dimension and achieved 69.8 percent on BrowseComp and 71.5 percent on BrowseComp-ZH, and MiroThinker-v1.0, which supported up to 600 tool calls per task across 8B, 30B, and 72B parameter variants.

The framework is optimized for financial prediction and general research tasks, with performance tracked across multiple benchmarks including HLE-Text, BrowseComp, BrowseComp-ZH, GAIA-Val-165, XBench-DeepSearch, and FutureX. The repository includes a live demonstration accessible at dr.miromind.ai, which supports research report generation, document preview and sharing, and accepts multiple file formats including PDF, DOC, PPT, XLS, and JPG uploads. GitGenius data shows the most active issue labels are enhancement requests with three tracked items and questions with one tracked item, indicating ongoing development and user engagement.

The project maintains connections with related repositories through overlapping contributors, including IBM's AssetOpsBench, Alibaba NLP's DeepResearch, and the vLLM project, suggesting integration with broader AI infrastructure and research communities. The repository includes comprehensive documentation covering quick start guides, benchmark evaluation procedures, trace collection methodologies, and FAQ sections to support both researchers and practitioners implementing the framework.