MiroThinker: Open-Source Deep Research Agent with 256K Context and 300 Tool Calls
MiroThinker is a deep research agent optimized for complex research and prediction tasks. Our latest models, MiroThinker-1.7, achieves 74.0 and 75.3 on the BrowseComp and BrowseComp Zh, respectively.
At a glance
- What is it?
- MiroThinker is an open-source deep research agent from MiroMindAI, released in 30B and 235B parameter scales. The 1.7 release achieves 74.0% on BrowseComp and 75.3% on BrowseComp-ZH, placing it among the leading open-source research agents on both benchmarks.
- Who is it for?
- Researchers and engineers who need an open-weight deep research agent with a large context window and published benchmark results should evaluate MiroThinker-1.7-mini (30B) as the entry point. Teams without GPU capacity for 30B models must use the hosted service at dr.miromind.ai.
- Can I use it commercially?
- Yes. Apache-2.0 is a permissive licence: you can use, modify and sell software built on it, as long as you keep its copyright and licence notices.
- Is it still maintained?
- Yes. The repository last received commits 86 days ago.
- What is it written in?
- Mainly Python, according to GitHub's language statistics.
Answers come from the project's GitHub data, last synced on September 30, 2026, and from our analysis. They are not legal advice.
Editorial analysis
What MiroThinker Is and Who It Is For
MiroThinker is a deep research agent designed for multi-step information gathering and prediction tasks. The project comes from MiroMindAI and provides open-weight model files alongside infrastructure code for running benchmarks and collecting traces. The models are released on HuggingFace under the miromind-ai organization.
The primary audience is researchers and engineers who need an agent that can perform long-horizon research: queries that require issuing many tool calls, synthesizing information from multiple sources, and maintaining coherent reasoning over a 256K token context window. The 1.7 series supports up to 300 tool calls per task. Practical use cases include financial prediction tasks, deep web research, and multi-step analytical questions that a single LLM call cannot address.
The repository also targets machine learning researchers who want to reproduce benchmark results or contribute new model evaluations to the project's test suite.
Interactive Scaling: The Third Dimension of Performance
The v1.5 README section introduces what the project calls interactive scaling. Traditional scaling approaches in language models focus on two dimensions: model size (number of parameters) and context length (tokens in the window). MiroThinker adds a third dimension by training the agent to handle more frequent and deeper interactions with its environment, specifically tool calls.
The MiroThinker-v1.0 release supported up to 600 tool calls per task. The 1.7 series caps at 300, prioritizing accuracy of stepwise reasoning over raw call count. The project characterizes MiroThinker-H1, its proprietary agent, as focused on long-chain verifiable reasoning: reasoning that is step-verifiable and globally verifiable. This is a design goal for complex tasks where you need to audit why the agent reached a particular conclusion, not just accept the final answer.
The 235B model in the 1.7 series achieves 74.0% on BrowseComp, 75.3% on BrowseComp-ZH, 82.7% on GAIA-Val-165, and 42.9% on HLE-Text. The 30B model (MiroThinker-1.7-mini) achieves 72.3% on BrowseComp-ZH.
Getting MiroThinker Running: HuggingFace and the Gradio Demo
The model weights for MiroThinker-1.7-mini and MiroThinker-1.7 are available at huggingface.co/miromind-ai. The repository contains a apps/ directory and a libs/ directory for tooling and workflows. A Gradio demo application for local experimentation was introduced in August 2025.
For users who do not have the hardware to run 30B or 235B models locally, MiroMindAI operates a hosted version at dr.miromind.ai. The online version supports report generation, preview, and sharing, as well as document uploads in formats including .pdf, .doc, .ppt, .xls, and .jpg.
The repository uses a justfile for common development tasks:
just lintThis runs ruff for linting. The justfile also defines sort-imports, format, and precommit targets. License compliance is enforced through the reuse tool.
Benchmark Performance Across MiroThinker Versions
MiroThinker has released several model versions since August 2025, each with published benchmark scores. The v0.2 release in September 2025 achieved 17.2% on BrowseComp-EN and 29.4% on BrowseComp-ZH. The v1.0 release in November 2025 improved to 47.1% on BrowseComp and 55.6% on BrowseComp-ZH, using a 256K context window and up to 600 tool calls per task.
The v1.5 series introduced a 235B model that scored 69.8% on BrowseComp and 71.5% on BrowseComp-ZH. The 30B model in the same series (MiroThinker-v1.5-30B) was described as surpassing Kimi-K2-Thinking on BrowseComp-ZH at lower cost.
The 1.7 series released in March 2026 achieved the current best published scores: 74.0% on BrowseComp and 75.3% on BrowseComp-ZH for the 235B model, and 72.3% on BrowseComp-ZH for the 30B mini model. The project describes the 235B 1.7 model as achieving state-of-the-art on BrowseComp-ZH among open-source models at the time of release.
These numbers are from the project's own benchmark reports. External independent reproduction of these results is not documented in the README.
Limitations: Compute Requirements, Closed Tasks, and Scope
The most immediate limitation is hardware. A 30B parameter model requires substantial GPU memory to run. The 235B model is not practical for individual researchers without access to a high-memory multi-GPU system. Users without that hardware must rely on the hosted service, which is controlled by MiroMindAI and may change its availability or pricing independently of the open-source repository.
MiroThinker is designed for research and prediction tasks that require web browsing and multi-step reasoning. It is not designed as a general-purpose chat model or a code generation assistant. Applying it to tasks outside its training focus may produce worse results than simpler models better suited to those tasks.
The benchmark evaluations published by the project focus on BrowseComp, BrowseComp-ZH, GAIA, and HLE. Performance on other task types is not covered in the README. The MiroVerse-v0.1 dataset, hosted on HuggingFace, is the training data published alongside the models, but the README does not describe its composition in detail.
The repository has no GitHub releases. The last push was on 2026-07-06.
MiroThinker vs. Perplexity: Open Model vs. Hosted Research Service
Perplexity is a hosted AI search and research service. It takes a query, retrieves relevant web pages, and synthesizes an answer with citations. It is a consumer-facing product with a subscription model, not an open-weight model that can be self-hosted.
MiroThinker is an open-weight model that can be run locally or through the MiroMindAI hosted service. The key difference for engineers is access and control. With MiroThinker, you can load the model weights, run your own benchmark evaluations, and integrate the model into a custom pipeline. With Perplexity, you interact through an API or web interface and have no control over the underlying model or its tool-call behavior.
The BrowseComp benchmark tests multi-hop web research questions where the answer requires synthesizing information from several pages. MiroThinker's benchmark numbers on BrowseComp are a direct measure of its research quality on that benchmark. Perplexity does not publish BrowseComp results for direct comparison.
Maintenance and License
The repository is not archived. The last push was on 2026-07-06. The project is licensed under Apache-2.0, which allows commercial use, modification, and distribution with attribution and license retention. The model weights released on HuggingFace may carry separate license terms that apply to model use and distribution, which are governed by the HuggingFace model cards rather than the repository's Apache-2.0 license.
The project operates a Discord server and an official website at miromind.ai. The README indicates the team intends to continue iterating on the models and supports external contributions to the benchmark evaluation tools.
Editorial conclusion
Researchers and engineers who need an open-weight deep research agent with a large context window and published benchmark results should evaluate MiroThinker-1.7-mini (30B) as the entry point. Teams without GPU capacity for 30B models must use the hosted service at dr.miromind.ai. The last push to the repository was on 2026-07-06, and the project is licensed under Apache-2.0.
Frequently asked questions
What is MiroThinker?
MiroThinker is an open-source deep research agent from MiroMindAI, optimized for complex multi-step research and prediction tasks. It is available in 30B (MiroThinker-1.7-mini) and 235B (MiroThinker-1.7) parameter scales on HuggingFace, with a 256K context window and up to 300 tool calls per task.
What is MiroMind AI?
MiroMind AI is the organization behind MiroThinker. They develop and release open-source deep research agents alongside a hosted research service at dr.miromind.ai, and they publish benchmark evaluations on BrowseComp, GAIA, and HLE.
How does MiroThinker differ from a standard RAG pipeline?
A standard RAG pipeline retrieves documents once and passes them to the model in a single call. MiroThinker issues up to 300 tool calls per task in an iterative loop, using each call's result to inform the next query, which allows it to handle research tasks that require navigating through multiple levels of information rather than a single retrieval step.
Official sources
Add this badge to your README
If you maintain this project, the badge below links readers to this analysis and shows its maintenance status from the daily GitHub snapshot. Paste the markdown into your README; add ?metric=license or ?metric=stars to the image URL for a different field.
[](https://hysenlabs.com/projects/miromindai-mirothinker)