DOCSEARCHbyAlgolia
Submit Docs

MiroThinker Documentation

repository·main·Indexed 27 days ago

https://github.com/miromindai/mirothinker

A high-performance deep research agent optimized for research and prediction tasks, featuring specialized versions for financial prediction and complex web browsing. The ecosystem includes MiroThinker-v1.5-30B, the MiroFlow Agent for executing research tasks and benchmarks (such as BrowseComp and HLE), and the collect-trace package for capturing multi-turn interaction traces to create SFT/DPO training data.

Tokens
24.7K
Snippets
62
Records
148
Agent score
93%

What's inside MiroThinker

  1. Overview of the Collect Trace workflow

    main

    The collect-trace component automates the creation of training data by following these steps:

    1. Load RLVR items: Reads questions and verifiable answers from a dataset.
    2. Run Agent: Executes the agent in an evaluation pipeline (including tool use/browsing).
    3. Verify: Uses an LLM-as-a-Judge to compare the model's answer against the RLVR reference.
    4. Harvest: For correct items only, collects the full multi-turn trace and converts it into SFT/DPO-ready samples.

  2. Overview of MiroThinker-1.7 Models

    main
    MiroThinker-1.7 is a family of models designed for reliable agents performing long-chain tasks and deep research. They feature a 256K context window, support up to 300 tool interactions per task, and utilize enhanced post-training for stepwise reasoning. The family includes 30B and 235B parameter scales.

  3. Overview of MiroThinker-v1.5 Models

    main
    MiroThinker v1.5 is an open-source search agent that uses interactive scaling to handle deeper agent-environment interactions. It supports a 256K context window and can handle up to 400 tool calls per task. It is available in 30B and 235B parameter scales.

  4. Overview of MiroThinker-v1.0 Models

    main
    MiroThinker v1.0 introduces interactive scaling at the agent level, training agents to handle frequent agent-environment interactions using feedback to correct errors. It supports a 256K context window and up to 600 tool calls per task. Models are available in 8B, 30B, and 72B scales.

  5. Overview of MiroThinker

    main
    MiroThinker is a deep research agent optimized for research and prediction tasks. It is designed to perform complex information retrieval and reasoning, achieving high performance on benchmarks like BrowseComp. The project provides various model scales (e.g., 8B, 30B, 72B, 235B) and specialized versions optimized for different domains such as financial prediction.

  6. Overview of MiroThinker-v0.2 Models

    main
    MiroThinker v0.2 features richer training data (English and Chinese), unified DPO training, and an extended context length of 64K. It is available in SFT and DPO variants across 4B, 8B, 14B, and 32B scales.

  7. Overview of MiroThinker-v0.1 Models

    main
    MiroThinker v0.1 includes SFT and DPO variants at 8B, 14B, and 32B scales, based on Qwen3. It features a 40K context window and achieved state-of-the-art performance on the GAIA benchmark among open-source models.

  8. Key Features of the MiroThinker-Optimized Framework

    main

    The MiroThinker-Optimized Framework provides several core capabilities for agent development:

    • Fully Open-Source Agent Framework: Offers complete transparency through open framework and open agents.
    • Tool Integration: Supports seamless integration with external tools and APIs.
    • Trace Collection: Provides comprehensive logging and analysis of agent interactions. Traces include elapsed time and estimated completion time (displayed in minutes), making the data ready for SFT (Supervised Fine-Tuning) and DPO (Direct Preference Optimization).
    • Benchmark Evaluation: Includes extensive testing capabilities across multiple benchmark datasets.

  9. Hardware Requirements for MiroThinker v1.5

    main

    To run the MiroThinker v1.5-30B (a 30B MoE model with 3B active parameters) locally, ensure your hardware meets these specifications:

    • GPU: NVIDIA RTX 40xx/50xx series or equivalent.
    • VRAM:
      • 16GB minimum: Required if using Q4 quantization via llama.cpp.
      • 48GB+ recommended: Required for FP8 quantization or handling longer context windows.

  10. Deploy LLM Server using SGLang (Recommended)

    main

    SGLang provides high-performance inference using FP8 quantization. This method starts an OpenAI-compatible server at http://0.0.0.0:61005/v1.

    Prerequisites:

    • Install SGLang.
    • Download the model checkpoints as described in the checkpoint guide.
    MODEL_PATH=model/MiroThinker-v1.5-30B

python3 -m sglang.launch_server \
    --model-path $MODEL_PATH \
    --mem-fraction-static 0.9 \
    --quantization fp8 \
    --tp 1 \
    --dp 1 \
    --host 0.0.0.0 \
    --port 61005 \
    --trust-remote-code

  11. Serve MiroThinker Models with SGLang

    main

    To serve MiroThinker models at a specific port (e.g., 61002) using SGLang:

    NUM_GPUS=4
PORT=61002
AGENT_PATH=miromind-ai/MiroThinker-1.7-mini

python3 -m sglang.launch_server \
    --model-path $AGENT_PATH \
    --tp $NUM_GPUS \
    --dp 1 \
    --host 0.0.0.0 \
    --port $PORT \
    --trust-remote-code

    The server will be available at http://0.0.0.0:$PORT. Use this as your llm.base_url (e.g., http://0.0.0.0:61002/v1).

    NUM_GPUS=4
PORT=61002

# Downloading agent from HF 
AGENT_PATH=miromind-ai/MiroThinker-1.7-mini


python3 -m sglang.launch_server \
    --model-path $AGENT_PATH \
    --tp $NUM_GPUS \
    --dp 1 \
    --host 0.0.0.0 \
    --port $PORT \
    --trust-remote-code

  12. Collect Traces for SFT and DPO

    main

    Navigate to apps/collect-trace to run scripts for collecting traces used in Supervised Fine-Tuning (SFT) or Direct Preference Optimization (DPO).

    cd apps/collect-trace

# Collect Traces for SFT
bash scripts/collect_trace_claude37.sh
bash scripts/collect_trace_gpt5.sh

# Collect Traces for DPO
bash scripts/collect_trace_qwen3.sh