MiroThinker is an open-source deep research agent and model series from MiroMind, published on GitHub with around 8,400 stars. Built for complex research and prediction tasks, it browses, searches, and reasons over many steps to produce grounded, evidence-backed answers rather than single-shot replies.
Deep research tasks demand long chains of tool use: searching the web, opening sources, cross-checking claims, and revising conclusions. MiroThinker is optimized for exactly that, sustaining reasoning across 100+ interaction steps. The project reports strong benchmark results for an open system โ 74.0 on BrowseComp and 75.3 on BrowseComp-ZH for the latest MiroThinker-1.7 release, and a 60.2% Avg@8 score on the GAIA benchmark โ alongside competitive results on other evaluations such as HLE and FutureX. The release includes both the agent framework and the open model checkpoints in several sizes, so teams can self-host the full stack or use the agent harness with the provided models. Because everything is published openly, researchers can inspect how the agent plans searches, manages context over long horizons, and assembles its final report.
MiroThinker is free and open source under the Apache-2.0 license. The models and agent code are on GitHub; there is no paid hosted tier, so your only cost is the compute you run it on.
MiroThinker suits research teams and agent developers who need an open deep-research system they can self-host, audit, and fine-tune โ for example for market research, literature review, forecasting, and competitive intelligence workloads.