November 13, 2025 marks an exciting milestone in AI research—the MiroMind AI team officially released MiroThinker v1.0, an open-source research agent. This isn't just another large language model release; it represents a groundbreaking new approach to enhancing AI agent capabilities.
🎯 Revolutionary Innovation: The Third Dimension of Scaling
Throughout AI's evolution, we've witnessed two primary dimensions of performance improvement: model parameter scale and context window length. MiroThinker v1.0 boldly introduces a third dimension—Interactive Scaling.
What is Interactive Scaling?
Interactive scaling is a systematic training methodology that teaches models to engage in deeper and more frequent interactions with their environment. Through environment feedback and external information acquisition, models can proactively correct errors and refine reasoning trajectories, significantly boosting task completion capabilities.
Think of it like training a researcher. It's not simply about having them memorize more knowledge (parameter scale) or read longer documents at once (context length), but rather teaching them how to efficiently use tools, search for information, verify facts, and continuously adjust their research strategy during exploration.
The experimental data is compelling: RL-tuned MiroThinker v1.0-30B exhibits far longer and deeper interaction trajectories than its SFT counterpart. Across all four major benchmarks, while supervised fine-tuning models terminate after just a few tool calls, the RL model performs extended multi-turn reasoning, exploration, and information verification. This behavioral shift yields 8-10 percentage point accuracy gains, clearly demonstrating the positive correlation between interaction depth and performance.
💪 Impressive Technical Specifications
MiroThinker v1.0 showcases remarkable technical capabilities:
Core Parameters
- Context Window: 256K tokens (supporting ultra-long document processing and deep analysis)
- Tool Call Capability: Up to 600 tool calls per task—a substantial breakthrough among open-source research agents
- Multi-Scale Release: Available in 8B, 30B, and 72B parameter versions to accommodate different computational budgets and application scenarios
| Model Name | Base Model | Max Length | Max Tool Calls |
|---|---|---|---|
| MiroThinker-v1.0-8B | Qwen3-8B | 256K | 600 |
| MiroThinker-v1.0-30B | Qwen3-30B-A3B-Thinking-2507 | 256K | 600 |
| MiroThinker-v1.0-72B | Qwen2.5-72B-Instruct | 256K | 600 |
Why are 600 Tool Calls Important?
For complex research tasks, agents need to:
- Search multiple information sources
- Cross-verify data accuracy
- Deep-dive into relevant literature
- Synthesize analysis and draw conclusions
Support for 600 tool calls enables MiroThinker to conduct deep, multi-step information exploration and verification like a genuine researcher, without interruption due to call limitations.
📊 Outstanding Benchmark Performance
MiroThinker v1.0 demonstrates strong performance across multiple authoritative benchmarks:
- HLE-Text (Humanity's Last Exam - Text): 37.7%
- BrowseComp: 47.1% (approaching DeepResearch's 51.5%)
- BrowseComp-ZH (Chinese Browse Comprehension): 55.6%
- GAIA-Text-103: 81.9%
These results not only surpass previous open-source agents but, more importantly, significantly narrow the gap with commercial models like GPT-5-high. Notably, on the HLE-Text benchmark, MiroThinker even outperforms GPT-5-high by 2.5 percentage points!
🛠️ Real-World Application Scenarios
MiroThinker v1.0's capabilities make it ideal for numerous use cases:
1. Academic Research Assistant
- In-depth literature reviews and information synthesis
- Cross-disciplinary knowledge correlation analysis
- Research hypothesis verification and data collection
2. Business Intelligence Analysis
- Market trend research and competitive analysis
- Multi-source information cross-verification
- Strategic decision support
3. Technical Problem Solving
- Complex technical documentation comprehension
- Multi-step problem diagnosis
- Solution exploration and evaluation
4. Content Creation Research
- Deep background investigation
- Fact-checking and citation verification
- Multi-perspective information integration
🌟 The Power of Open Source
MiroThinker v1.0 is released under the MIT License, which means:
- ✅ Completely Free to Use: For both commercial and non-commercial projects
- ✅ Full Code Access: Learn, modify, and customize as needed
- ✅ Community-Driven Development: Contributions and collaboration welcome
- ✅ Transparent and Trustworthy: All technical details publicly available
This open approach not only democratizes AI technology but also provides invaluable learning and innovation opportunities for the research community.
🚀 Quick Start
Want to experience MiroThinker v1.0's powerful capabilities? Multiple options available:
Online Demo
Visit the official demo site: https://dr.miromind.ai/
Local Deployment
Access the GitHub repository for complete installation instructions:
👉 https://github.com/MiroMindAI/MiroThinker
Hugging Face Models
🔮 Looking Ahead
The release of MiroThinker v1.0 marks an important milestone in AI agent development. The introduction of interactive scaling as a third dimension of performance improvement points us toward more general and intelligent AI systems.
With the synergistic optimization of model scale, context length, and interaction depth, we have every reason to believe that future AI agents will handle increasingly complex, open-ended tasks, truly becoming powerful assistants in human research and creation.
The MiroMind AI team commits to continued R&D investment, continuously improving the MiroThinker series, and looks forward to collaborating with developers and researchers worldwide to advance this technology.
📞 Contact Information
If you have any questions or suggestions about MiroThinker v1.0, feel free to contact the MiroMind AI team through:
- 📧 Email: service@miromind.ai
- 💬 GitHub: https://github.com/MiroMindAI/
- 🎮 Discord: Join the Community
- 📱 WeChat: Search for MiroMind official account
- 📝 RedNote: Follow MiroMind
Reference Resources:
This is not just a model release—it's a pivotal step in the evolution of AI agents. Let's witness together how interactive scaling redefines the future of AI! 🎉
Top comments (0)