MiroThinker
Open-weight deep research agent models with a tool-using search framework
What it is
MiroThinker is a family of open-weight language models trained to act as research agents, searching the web and calling tools across hundreds of steps to answer hard questions and make predictions. Releases come in 30B and 235B sizes with a 256K context window, alongside code for running the agent, a Gradio demo, benchmark evaluation on tasks like BrowseComp, GAIA and HLE, and trace collection. Researchers and developers use it to run or study long-horizon search agents on their own hardware.
