MiroThinker
a deep-research agent team that verifies its own evidence
Last verified:
What is MiroThinker?
MiroThinker is a series of open-source state-of-the-art Deep Research models designed for reasoning, not just conversation. It features a 235B parameter reasoning model built with a Verification-Centric Architecture that enables users to predict, verify, and discover with science-based AI. The tool excels in complex multi-step reasoning tasks, supporting up to 100+ step DAG reasoning under MiroMind OS with stability and accuracy.
Key features include a massive 256K context window, capability for up to 600 rounds of tool interactions per task, and Interactive Scaling technology that trains agents to handle deeper and more frequent environment interactions. MiroThinker achieves exceptional benchmark performance with 60.2% Avg@8 on GAIA, 47.1% on BrowseComp (close to OpenAI DeepResearch's 51.5%), and 37.7% on HLE human eval (above GPT-5-high with Python + Search). The model is built on Qwen-3 base and supports tool-augmented reasoning for complex research workflows.
MiroThinker is designed for researchers, AI developers, data scientists, and anyone needing deep information seeking capabilities. It serves users working on complex research tasks, scientific analysis, prediction workflows, coding agents, and multi-step analytical problems. The fully open-source model and code are accessible via GitHub and Hugging Face, making it ideal for those who want transparency and customization in their AI research agents.
MiroThinker pricing
Pricing model: Freemium
MiroThinker uses a credit-based pricing system with dynamic consumption calculated based on model token usage and tool invocation costs. Three tiers exist: Pro Mode (S: 200 credits, M: 600 credits, L: 1,600 credits) and Standard Mode at 50% of Pro (S: 100 credits, M: 300 credits, L: 800 credits). Users receive 100 free credits daily, enough for one Standard mode query. Credit consumption varies by question complexity, model usage, and tool calls. Most routine questions fall into S tier. Regular credit giveaways and promotions available for active users. The model itself is 100% fully open-source with free access to model and code on GitHub and Hugging Face.
MiroThinker pros
- Fully open-source model and code with complete transparency
- 235B parameter reasoning model for maximum performance
- 256K context window for handling very long documents
- Up to 600 rounds of tool interactions per task
- 60.2% Avg@8 score on challenging GAIA benchmark
- 47.1% BrowseComp accuracy接近OpenAI DeepResearch
- 37.7% HLE human eval above GPT-5-high with Python + Search
- Verification-Centric Architecture for accurate reasoning
- Interactive Scaling technology for deeper environment interactions
- Supports 100+ step DAG reasoning under MiroMind OS
- Built on Qwen-3 base model for strong foundation
- Temporal-Sensitive Training for causality mastery
- Predictive capabilities for forward-looking analysis
- Multiple model sizes available (8B, 30B, 72B, 235B)
- 30B model costs only 1/20 of Kimi-K2 for efficiency
- MCP server support with flexible configuration options
- Self-correction capabilities for coding errors
- Robust guardrails and safety alignment
- Available via interactive demo at dr.miromind.ai
- API access through OpenAI-compatible endpoint
MiroThinker cons
- Credit consumption dynamically calculated, not fixed per request
- Requires only 3 MCP servers but setup complexity for beginners
- Pro Mode credits expensive (1,600 for Complex tasks)
- Only 100 free credits daily, limited for heavy usage
- 235B model requires significant computational resources
- Best configurations like keep5 require technical knowledge
- No traditional subscription tiers, only credit-based pricing
- Tool invocation costs add to overall credit consumption
- Complex research tasks may exhaust daily credits quickly
- Limited documentation for non-technical users
Frequently asked questions about MiroThinker
What is MiroThinker?
MiroThinker is a series of open-source state-of-the-art Deep Research models designed for reasoning, not just conversation. It features a 235B parameter reasoning model with a Verification-Centric Architecture that enables prediction, verification, and discovery with science-based AI.
Is MiroThinker open-source?
Yes, MiroThinker is 100% fully open-source. Both the model and its code are fully accessible and available on GitHub and Hugging Face, providing free access to discovery-driven intelligence.
What benchmarks does MiroThinker achieve?
MiroThinker achieves 60.2% Avg@8 on GAIA benchmark, 47.1% on BrowseComp (close to OpenAI DeepResearch's 51.5%), and 37.7% on HLE human eval (above GPT-5-high with Python + Search at 35.2%). The v1.5 version achieves 80.8% on GAIA-Val-165.
What is the context window size?
MiroThinker has a 256K context window, enabling it to handle very long documents and maintain context across extended research workflows.
How many tool interactions can MiroThinker perform?
MiroThinker supports up to 600 rounds of tool interactions per task, a dramatic increase from previous standards, enabling sustained multi-turn reasoning for complex research workflows.
What is Interactive Scaling?
Interactive Scaling is MiroThinker's third dimension of AI scaling that trains models to engage in deeper, more frequent interactions with external tools and environments. Instead of just outputting text, it iteratively searches, validates findings, and refines reasoning based on environmental feedback.
What pricing model does MiroThinker use?
MiroThinker uses a credit-based pricing system with dynamic consumption. Pro Mode has S (200), M (600), L (1,600) credit tiers. Standard Mode is 50% of Pro. Users get 100 free credits daily for one Standard mode query.
What model sizes are available?
MiroThinker is available in multiple sizes: 8B, 30B, 72B, and 235B parameter variants. The 30B model offers the best balance of performance and efficiency, while the 235B provides maximum performance for critical tasks.
Can I access MiroThinker via API?
Yes, MiroThinker is accessible via API through an OpenAI-compatible endpoint at api.miromind.ai/v1/chat/completions. Users need to login to the console and create an API key from the API Keys page.
What is MiroMind OS?
MiroMind OS is the operating system for mirroring human intelligence and awakening AI self-awareness through the continuity of memory. It provides the foundation for MiroThinker's stability and accuracy in complex reasoning tasks.