MiroThinker v1.5 is a top-notch open-source AI search agent with a 256K context window and up to 400 tool calls in a single task, enabling deep network research, code execution, and multi-step reasoning. The tool topped benchmarks such as HLE-Text (39.2%), BrowseComp (69.8%), and GAIA (80.8%), surpassing other free agents with a low cost advantage. You can get accurate real-world research support – such as finding arXiv academic papers, answering complex queries, etc., which is more efficient and less expensive than paid tools, and you can get full open source access to the tool on GitHub and Hugging Face.
It is an open-source AI agent designed for “deep search and research”.
Positioning MiroThinker
MiroThinker v1.5 is a high-performance, low-cost, in-depth AI search agent.
It is specially used for: checking data, running code, and multi-step reasoning.
It comes from MiroMindAI,
The objectives are very clear:
In the matter of “search + reasoning”, it is stronger than most free agents,
At the same time, it does not rely on burning money.
How is it different from regular AI search tools?
Let’s take a look at three hard indicators first, which is also the most “hardcore” part of MiroThinker.
256K Context Window: It’s productivity
What does 256K context mean?
It’s not “we can talk longer”,
Instead:
- It can eat a large amount of web content at once
- Able to integrate multiple papers, reports, and code
- Eliminate the need to lose context repeatedly and reduce information loss
For research tasks, this is a qualitative change.
You can make it:
- More than a dozen arXiv papers were analyzed simultaneously
- Compare the details of multiple technical solutions
- Long-link reasoning without “amnesia”
Up to 400 tool calls in a single task: a true “can run”
Many agents claim to “support tool calls”,
But it was actually only called a few times before it ended.
MiroThinker v1.5 is designed for “long tasks”:
- Search → Parse → and then search
- Download → Run → Analysis → Fix
- Failure in the middle can continue
400 tool calls mean:
It is not a one-time performance, but it works continuously.
This is also the key to its ability to perform “in-depth network research”.
Not just searching, but also running code
MiroThinker does more than just “check”.
It can:
- Execute code
- Processing data
- Perform secondary calculations and verification of search results
This is very important in complex issues.
For example:
You don’t just want to “find the answer”,
Instead, I want to confirm whether this answer is reliable.
What do benchmark test results say?
Many projects avoid benchmarking,
But MiroThinker is the opposite.
In several very practical evaluation sets, its results are very eye-catching:
- HLE-Text：39.2%
- BrowseComp：69.8%
- GAIA：80.8%
These tests are essentially not about “whether you can chat”,
Instead:
Can you check the information?
Can you reason clearly?
Whether the task can be completed in a complex environment
In these dimensions, MiroThinker has stablely surpassed most free agents.
“Cost-effective”?
Here is a big truth.
A lot of “strong-looking” AI Agents:
- or rely on expensive models
- or consume a lot of tokens
- or not open source at all
One of the core strengths of MiroThinker is:
In the free/low-cost range, performance is pushed to the limit.
This is crucial for individual developers, students, and independent researchers.
Github：https://github.com/MiroMindAI/MiroThinker
Tubing: