MiroThinker

• [2025-11-13] 🎉 MiroThinker-v1.0 is now released! Introducing interactive scaling as a third dimension of performance improvement, MiroThinker v1.0 supports 256K context window and up to 600 tool calls per task. Available in 8B, 30B, and 72B parameter scales, achieving 37.7%, 47.1%, 55.6%, and 81.9% on HLE-Text, BrowseComp, BrowseComp-ZH, and GAIA-Text-103, respectively. See Technical Report for more details.
• [2025-09-11] MiroThinker-72B-Preview ranked 4th in this week's FutureX benchmark. See FutureX.
• [2025-09-08] MiroThinker-v0.2 is now released, achieving open-source SOTA performance across multiple benchmarks, including HLE (17.8%), HLE-Text-Only (19.1%), BrowseComp-EN (17.2%), BrowseComp-ZH (29.4%), XBench-DeepSearch (56.0%), and Frames (74.8%).
• [2025-09-07] We supported more benchmarks, including BrowseComp-ZH, XBench-DeepSearch, and FutureX. We plan to add more benchmarks in the future.
• [2025-08-22] Introducing streamlined deployment options for MiroThinker with optimized resource usage and faster startup times. Experience the interactive demo: 🚀 Try Gradio Demo
• [2025-08-08] MiroThinker-v0.1 released.

MiroThinker v1.5 is the world-leading open-source search agent that advances tool-augmented reasoning through interactive scaling — training the agent to handle deeper and more frequent agent-environment interactions as a third dimension of performance improvement, beyond model size and context length.

Key Features

• 🚀 MiroThinker v1.5 supports a 256K context window, long-horizon reasoning, and deep multi-step analysis.
• 🔧 Handles up to 400 tool calls per task — a substantial improvement over previous open-source research agents.
• 📦 Released in 30B and 235B parameter scales, accompanied by a comprehensive suite of tools and workflows to flexibly support diverse research settings and compute budgets.

MiroThinker v1.5 demonstrates strong general-research performance across a broad range of benchmarks, achieving 39.2%, 69.8%, 71.5%, and 80.8% on HLE-Text, BrowseComp, BrowseComp-ZH, and GAIA-Val-165, respectively. These results surpass previous open-source agents and set the new world-leading BrowseComp performance.

Unlike previous agents that scale only model size or context length, MiroThinker v1.0 introduces interactive scaling at the agent level, systematically training the agent to handle deeper and more frequent agent–environment interactions as a third dimension of performance improvement. Interactive scaling leverages environment feedback and external information acquisition to correct errors and refine trajectories.

• 🚀 256K Context Window: Supports long-horizon reasoning and deep multi-step analysis
• 🔧 600 Tool Calls: Handles up to 600 tool calls per task — a substantial improvement over previous open-source research agents
• 📦 Multiple Scales: Released in 8B, 30B, and 72B parameter scales, accompanied by a comprehensive suite of tools and workflows to flexibly support diverse research settings and compute budgets

MiroThinker v1.0 demonstrates strong general-research performance across a broad range of benchmarks, achieving 37.7%, 47.1%, 55.6%, and 81.9% on HLE-Text, BrowseComp, BrowseComp-ZH, and GAIA-Text-103, respectively. These results surpass previous open-source agents and narrow the gap with commercial counterparts such as GPT-5-high.

In this new version, we introduced three key improvements:

• 📚 Richer training data from both English and Chinese sources, yielding significant gains in benchmark performance and generalization
• 🎯 Unified DPO training with a single preference dataset across all agents
• 📏 Extended context length from 40k to 64k for more challenging multi-turn tool-use tasks

Compared to v0.1, MiroThinker v0.2 delivers consistent gains across benchmarks. For example, scores improved from 57.3 → 64.1 on GAIA-Text-103 and from 17.0 → 29.4 on BrowseComp-ZH, reflecting substantial advancements in the model’s general research agent capabilities.

Performance of Open-Source Agents on GAIA-Validation Benchmark.

We have released the MiroThinker v0.1 series, including both SFT and DPO variants at parameter scales of 8B, 14B, and 32B. Notably, MiroThinker v0.1 achieves state-of-the-art performance among open-source models on the GAIA benchmark, a rigorous evaluation suite for advanced agentic capabilities, demonstrating its strength in long-context, decision-intensive, and real-world task scenarios.

• GAIA Validation: A benchmark for General AI Assistants. (paper)
• GAIA-Text-103: A subset of GAIA Validation for text-only tasks. (paper)
• HLE: Humanity's Last Exam. (paper)
• HLE-Text-2158: A subset of HLE for text-only tasks. (paper)
• HLE-Text-500: A subset of HLE for text-only tasks, created by WebThinker. (paper)
• BrowseComp-EN: Web browsing and comprehension tasks. (paper)
• BrowseComp-ZH: A Chinese version of BrowseComp. (paper)
• WebWalkerQA: Web navigation and question answering. (paper)
• Frames: Factuality, Retrieval, And reasoning MEasurement Set. (paper)
• XBench-DeepSearch: A benchmark for deep research agents. (website)
• FutureX: A live benchmark designed for predicting unknown future. (website)
• SEAL-0: A benchmark for evaluating LLMs on conflicting-evidence web questions. (paper)
• AIME2025: American Invitational Mathematics Examination 2025. (website)
• DeepSearchQA: Google's Deep Search Question Answering benchmark. (paper)

To prevent potential information leakage (e.g., searching benchmark answers from HuggingFace), access to HuggingFace has been explicitly disabled in these tools.

We further perform canary string testing on the tool outputs of all trajectories and disregard any trajectory found to be contaminated, treating it as an incorrect answer.

1. Following the practices of WebThinker, WebAgents, and CognitiveKernel, we report the Best Pass@1, the highest score across three runs, which often reflects stronger performance, though it may exhibit some variability. To provide a more stable measure, we additionally report Pass@1 (Avg@8), which offers greater consistency at the cost of slightly lower scores.

2. For consistency with prior open-source works, we evaluate GAIA-Text-103 using the WebAgents LLM-as-a-Judge template, and report results on GAIA-Val-165 using the official GAIA scorer script.

3. By default, we use open-source tools wherever possible, except for the code tool E2B and the Google search tool Serper. We use Whisper, Qwen2.5-VL-72B-Instruct, and Qwen3-235B-A22B-Thinking-2507 in our implementation. The framework can be easily extended to other open-source tools of your choice.

4. Replacing these open-source tools with commercial alternatives can yield performance gains. Commercial tools were mainly used for multimodal capabilities and certain complex reasoning subtasks. The majority of tasks, including planning, browsing, refinement, navigation, and more, were handled by our agents.

1. MiroThinker’s performance was tested with this repository and open-source tools; other agents’ results are from their papers and official sites.

2. As MiroVerse-v0.1 mainly contains English data, the agent’s Chinese capability is limited. We plan to add more Chinese data to improve performance in the next version.

The following optional tools are available but were not used in MiroThinker v1.0-1.7 evaluation:

📖 Local Deployment: For instructions on deploying open-source tools (tool-vqa-os, tool-transcribe-os, tool-reasoning-os) locally, see Local Tool Deployment Guide.

See the MiroFlow Tools README for complete documentation of all available tools.

You can create your own YAML configuration file to freely combine MCP servers. Here's how:

1. Create a new YAML file in apps/miroflow-agent/conf/agent/:

# conf/agent/my_custom_config.yaml
defaults:
  - default
  - _self_

main_agent:
  tools:
    - tool-python                    # Execution environment
    - search_and_scrape_webpage      # Google search
    - jina_scrape_llm_summary        # Web scraping with LLM
    - tool-vqa                       # Vision processing (optional)
    - tool-transcribe                # Audio processing (optional)
    - tool-reasoning                 # Reasoning engine (optional)
    - tool-reading                   # Document reading (optional)
  max_turns: 300  # Maximum number of turns

sub_agents:
  agent-browsing:  # Optional sub-agent
    tools:
      - tool-google-search
      - tool-vqa
      - tool-reading
      - tool-python
    max_turns: 50

keep_tool_result: -1  # Context retention budget: -1 keeps all tool results, or specify K to keep only the K most recent tool responses

💡 Context Retention Strategy: The keep_tool_result parameter implements a recency-based context retention strategy. In the standard ReAct paradigm, all tool outputs are retained in the message history, which can lead to inefficient context utilization. Empirically, we observe that the agent's subsequent actions depend primarily on recent observations rather than distant ones. This strategy retains only the most recent K tool responses (where K is the keep_tool_result value) while preserving the complete sequence of thoughts and actions.

Benefits:

• ✅ Preserves the reasoning and action trace
• ✅ Focuses the agent's attention on the most contextually relevant observations
• ✅ Frees additional context space for extended reasoning and deeper tool-use trajectories
• ✅ Does not lead to performance degradation while allowing more context space for interactive scaling

Usage: Set keep_tool_result: -1 to keep all tool results, or specify a positive integer K (e.g., keep_tool_result: 5) to keep only the K most recent tool responses.

2. Use your custom configuration when running evaluations:

cd apps/miroflow-agent
uv run main.py llm=qwen-3 agent=my_custom_config llm.base_url=https://your_base_url/v1

3. Configure environment variables in .env based on the tools you use.

   All available environment variables are listed in apps/miroflow-agent/.env.example. Copy it to .env and configure the variables according to your chosen configuration:

   cd apps/miroflow-agent
   cp .env.example .env
   # Edit .env with your actual API keys

   For MiroThinker v1.5 (mirothinker_v1.5_keep5_max200.yaml, mirothinker_v1.5_keep5_max400.yaml, or mirothinker_v1.5.yaml) and v1.0 (mirothinker_v1.0_keep5.yaml or mirothinker_v1.0.yaml), see the Minimal Configuration section above for the complete configuration example.

   For other configurations, refer to the Pre-configured Agent Settings table above to see which environment variables are required.

⚠️ Important for MiroThinker-1.7: To reproduce our reported results, you must set the correct AGENT_SET:

• BrowseComp & BrowseComp-ZH: Use AGENT_SET="mirothinker_1.7_keep5_max300"
• All other benchmarks: Use AGENT_SET="mirothinker_1.7_keep5_max200"

# Navigate to the miroflow-agent directory first
cd apps/miroflow-agent

# HLE
NUM_RUNS=3 LLM_MODEL="xxx" BASE_URL="xxx" AGENT_SET="mirothinker_1.7_keep5_max200" bash scripts/run_evaluate_multiple_runs_hle.sh

# HLE-Text-2158
NUM_RUNS=3 LLM_MODEL="xxx" BASE_URL="xxx" AGENT_SET="mirothinker_1.7_keep5_max200" bash scripts/run_evaluate_multiple_runs_hle-text-2158.sh

# HLE-Text-500
NUM_RUNS=3 LLM_MODEL="xxx" BASE_URL="xxx" AGENT_SET="mirothinker_1.7_keep5_max200" bash scripts/run_evaluate_multiple_runs_hle-text-500.sh

# GAIA-Text-103
NUM_RUNS=8 LLM_MODEL="xxx" BASE_URL="xxx" AGENT_SET="mirothinker_1.7_keep5_max200" bash scripts/run_evaluate_multiple_runs_gaia-validation-text-103.sh

# GAIA-Validation (GAIA-Val-165)
NUM_RUNS=8 LLM_MODEL="xxx" BASE_URL="xxx" AGENT_SET="mirothinker_1.7_keep5_max200" bash scripts/run_evaluate_multiple_runs_gaia-validation.sh

# BrowseComp-EN (⚠️ use max300)
NUM_RUNS=3 LLM_MODEL="xxx" BASE_URL="xxx" AGENT_SET="mirothinker_1.7_keep5_max300" bash scripts/run_evaluate_multiple_runs_browsecomp.sh

# BrowseComp-ZH (⚠️ use max300)
NUM_RUNS=3 LLM_MODEL="xxx" BASE_URL="xxx" AGENT_SET="mirothinker_1.7_keep5_max300" bash scripts/run_evaluate_multiple_runs_browsecomp_zh.sh

# WebWalkerQA
NUM_RUNS=3 LLM_MODEL="xxx" BASE_URL="xxx" AGENT_SET="mirothinker_1.7_keep5_max200" bash scripts/run_evaluate_multiple_runs_webwalkerqa.sh

# XBench-DeepSearch
NUM_RUNS=8 LLM_MODEL="xxx" BASE_URL="xxx" AGENT_SET="mirothinker_1.7_keep5_max200" bash scripts/run_evaluate_multiple_runs_xbench_deepsearch.sh

# FRAMES
NUM_RUNS=3 LLM_MODEL="xxx" BASE_URL="xxx" AGENT_SET="mirothinker_1.7_keep5_max200" bash scripts/run_evaluate_multiple_runs_frames.sh

# SEAL-0
NUM_RUNS=8 LLM_MODEL="xxx" BASE_URL="xxx" AGENT_SET="mirothinker_1.7_keep5_max200" bash scripts/run_evaluate_multiple_runs_seal-0.sh

# FutureX
NUM_RUNS=8 LLM_MODEL="xxx" BASE_URL="xxx" AGENT_SET="mirothinker_1.7_keep5_max200" bash scripts/run_evaluate_multiple_runs_futurex.sh

# AIME2025
NUM_RUNS=32 LLM_MODEL="xxx" BASE_URL="xxx" AGENT_SET="mirothinker_1.7_keep5_max200" bash scripts/run_evaluate_multiple_runs_aime2025.sh

# DeepSearchQA
NUM_RUNS=3 LLM_MODEL="xxx" BASE_URL="xxx" AGENT_SET="mirothinker_1.7_keep5_max200" bash scripts/run_evaluate_multiple_runs_deepsearchqa.sh

A: We recommend MiroThinker-1.7 ⭐ with the minimal configuration:

• v1.7 ⭐: Latest version with 256K context, world-leading performance. Use config (with context management):
  • mirothinker_1.7_keep5_max200 (up to 200 turns, recommended for most tasks)
  • mirothinker_1.7_keep5_max300 (up to 300 turns, only used for BrowseComp and BrowseComp-ZH)

A: You need these keys for minimal setup:

• SERPER_API_KEY: Get from Serper.dev (Google search API)
• JINA_API_KEY: Get from Jina.ai (Web scraping)
• E2B_API_KEY: Get from E2B.dev (Code execution sandbox)
• SUMMARY_LLM_API_KEY: Your LLM API credentials (for content summarization). Can be a small model like Qwen3-14B or GPT-5-Nano—the choice has minimal impact on performance.
• OPENAI_API_KEY: Get from OpenAI (Required for benchmark evaluation, used for LLM-as-a-Judge)
• OPENAI_BASE_URL: Optional, defaults to https://api.openai.com/v1. Can be changed to use OpenAI-compatible APIs.

A: Common issues:

• Check base URL format: Should end with /v1 (e.g., https://your-api.com/v1)
• Verify API key: Ensure API_KEY is set correctly in environment or script
• Check server status: Make sure your server is running and accessible
• Network issues: Verify firewall/network settings allow connections

A: Troubleshooting steps:

1. Check working directory: Make sure you're in apps/miroflow-agent directory
2. Verify environment: Run uv sync to ensure dependencies are installed
3. Check .env file: Ensure all required environment variables are set
4. Review logs: Check logs/ directory for detailed error messages
5. Verify data path: Ensure benchmark data is downloaded and in correct location

A: Solutions:

• Reduce context length: Set MAX_CONTEXT_LENGTH to a smaller value (e.g., 131072 for 128K)
• Use context management with fewer turns:
  • For v1.5: Use mirothinker_1.7_keep5_max200 or mirothinker_1.7_keep5_max300 (with context management)
• Reduce concurrent tasks: Set MAX_CONCURRENT to a smaller number (e.g., 5)
• Use smaller agents:
  • For v1.5: Try 30B instead of 235B
  • For v1.0: Try 8B or 30B instead of 72B

A: Common fixes:

• E2B errors: Verify E2B_API_KEY is valid and account has credits
• Serper errors: Check SERPER_API_KEY and rate limits
• Jina errors: Verify JINA_API_KEY and JINA_BASE_URL are correct
• LLM summarization errors: Check SUMMARY_LLM_* variables and agent availability

A: Use the progress monitoring scripts:

cd apps/miroflow-agent
python benchmarks/check_progress/check_progress_<benchmark_name>.py /path/to/logs

The scripts show completion status, elapsed time, and estimated remaining time.