XNews

MiroThinker: open-source research agent models

MiroThinker: open-source research agent models

MiroMind released MiroThinker (v1.5) as an open-source research agent family including 235B and 30B versions, supporting very long context and high-frequency tool interactions for iterative research workflows.

Key insights

MiroThinker released as open-source research agent with large context windows

MiroMind open-sourced MiroThinker v1.5, offering 235B and 30B parameter variants, MIT license, and a 256K token context window. The project includes model weights, toolchains, and an open training corpus of 147k training samples. The team markets MiroThinker as a research-oriented agent capable of deep, iterative web browsing and multi-step tool use.

Interaction-scaling and tool-call capacity emphasized

MiroMind emphasizes 'interaction scaling'—improving capabilities by increasing depth and frequency of environment interactions. Community posts report single-task tool-call capacities: one post claims up to 400 tool calls, another claims up to 600 tool calls per task. The model natively supports web browsing, code execution, and multi-format document handling to form a think-act-observe loop.

Performance claims vs closed models and benchmarks

Community benchmarks and demonstrations claim MiroThinker reaches or approaches closed-model performance on BrowseComp, HLE, GAIA-Val and other research benchmarks, with specific claims of competitive results versus GPT and Claude in complex retrieval and reasoning tasks. Posts highlight strong Chinese-language performance on BrowseComp-ZH and xBench-DeepSearch.

Community uptake and low-cost inference promoted

Users reported MiroThinker being used for forecasting (e.g., Polymarket, markets) and research workflows via web demos. One post highlights cost-efficiency claims (30B variant cost ≈ 1/20 of Kimi-K2) and availability of web and GitHub resources for deployment.

Source posts