XNews

Open-source research agent: MiroThinker

Open-source research agent: MiroThinker

MiroThinker 1.5 positions itself as a high-capability open research agent. - MiroThinker claims strong performance with a 30B parameter model outperforming a 1T-class model in some tasks and lower inference cost (~$0.07 per run reported). - Features include automated search, hypothesis generation, evidence collection, verification loops, multi-round self-correction and up to 400 tool interactions per task. - Releases include a 235B-parameter variant with 256K context window in some reports; metrics cited: HLE-Text 39.2%, GAIA-Val 80.8%. - The project is open-source (MIT) with public training samples and vLLM/SGLang deployment support; community adoption examples include plugin recommendations and Polymarket-style forecasting tests.

Key insights

MiroThinker Promotes Open Research Agents with Deep Tooling and Long Context

Posts describe MiroThinker as an open-source research agent family claiming multi-round reasoning, self-directed inquiry and the ability to call up to 400 tools in a single task. Implementations vary by reported parameter counts; one report cites a 30B variant with high efficiency and another lists a 235B variant with a 256K context window and specific benchmark scores. The project publishes training samples and supports common deployment toolchains.

Community Uses MiroThinker for Research and Forecasting Experiments

Early adopters used MiroThinker to aggregate plugin recommendations, generate structured research outputs and run forecasting-style experiments (Polymarket/markets). The project’s open-source license and sample datasets support community reproducibility and rapid iteration.

Source posts