MiroThinker
MiroThinker is an open-source deep research agent optimised for research and prediction. The README says the project has several parts: MiroThinker itself (the research agent models, reported at 60.2% Avg@8 on the GAIA benchmark for an earlier release), MiroFlow, an agent framework for tool-use tasks with a reproducible GAIA score of 82.4%, and MiroVerse, a 147k-sample open training dataset on Hugging Face. A hosted demo is available.
Version 1.5, released in January 2026, is described as a search agent optimised for financial prediction that scales through deeper and more frequent agent-environment interaction. It supports a 256K context window and up to 400 tool calls per task, in 30B and 235B parameter sizes. The repository also provides a benchmark evaluation suite (GAIA, HLE, BrowseComp in English and Chinese, WebWalkerQA and others) and trace collection that logs agent interactions in a form ready for SFT and DPO training. Running it needs your own tool and model API keys, and the code is MIT licensed.
Features
Why choose it
Trade-offs
Maintenance check
- Last commit
- 2026-03-23
- Commits, last 90 days
- 0
- Releases, last 90 days
- 0
- Latest release
- —
- Open issues
- 5
- Recent issues with at least one comment
- 15/17
Oldest week on the left; the latest four weeks are orange. Source: GitHub commit activity.
Status rule: archived on GitHub = Archived; last commit within 90 days = Active; 91 to 365 days = Slowing; older = Stale. All figures come from the GitHub API on the date shown.
Most discussed open issues
- #176 还是1.7,好久没有升级了，这个repo也没有怎么更新了，是遇到什么问题了？3 comments · 2026-07-26
- #180 Deep research stuck at verification process been more than 24 hours now.0 comments · 2026-09-25
- #179 提问不输出结果，扣了费用没有积分到账，反馈无人回应0 comments · 2026-09-23
Issue titles are quoted from GitHub. Open the link to check whether a problem still applies to the version you use.
Who it fits
- You want open-weights research-agent models (30B and 235B, 256K context) trained specifically for up to 400 tool calls per task.
- You are researching or fine-tuning tool-use agents and want the MiroFlow harness, reproducible GAIA/HLE benchmarks, and the 147k-sample MiroVerse dataset.
- You do not have multi-GPU infrastructure to host 30B/235B models and just want an off-the-shelf research app using commercial APIs; use GPT Researcher.
- You want a document RAG platform for enterprise PDFs rather than an open-weights research model; use RAGFlow or MaxKB.
How it differs
vs GPT Researcher★ 30.3kMiroThinker provides open-weights 30B/235B agent models, a training dataset, and the MiroFlow benchmark framework. GPT Researcher is an application-level Python library and web UI that runs on your existing OpenAI/Claude/Ollama APIs.
Compatibility
Quick start
Use cases
Alternatives
Featured in these lists
Comments
No comments yet. Be the first!