Skip to content

MiroThinker

MiroThinker (4B/7B/14B/32B) is our suite of open-source agentic models, designed to work seamlessly with the MiroFlow framework. Our models are specifically built to handle complex, multi-tool tasks, leveraging the reproducible and robust foundation that MiroFlow provides.

By combining MiroFlow's reliable orchestration with MiroThinker's advanced reasoning capabilities, we offer a powerful, end-to-end solution for building high-performing, reproducible AI agents.

These models are a direct result of our extensive data collection efforts, utilizing MiroFlow to generate high-quality, post-training agent trace data. This unique approach enables MiroThinker to excel in planning, executing, and reasoning through complex multi-step tasks.

Deploying MiroThinker-32B with MiroFlow

This guide explains how to deploy the MiroThinker-32B-DPO-v0.2 model from Hugging Face and integrate it with MiroFlow.

Prerequisites

  • SGLang installed
  • Sufficient GPU memory for the model
  • MiroFlow repository set up

Step 1: Deploy Model with SGLang

Deploy the MiroThinker-32B model using SGLang:

SGLang Server Deployment
python3 -m sglang.launch_server \
    --model-path miromind-ai/MiroThinker-32B-DPO-v0.2 \
    --tp 8 \
    --dp 1 \
    --host 0.0.0.0 \
    --port 61005 \
    --trust-remote-code \
    --chat-template qwen3_nonthinking.jinja

Important Notes

  • Adjust the --tp (tensor parallelism) parameter to match your number of GPUs
  • Download the chat template from: qwen3_nonthinking.jinja
  • Ensure the port you used (in this case 61005) is available on your system

Step 2: Configure MiroFlow

Once the SGLang server is running, configure MiroFlow by adding the following to your .env file:

Environment Configuration
OAI_MIROTHINKER_API_KEY="dummy_key"
OAI_MIROTHINKER_BASE_URL="http://localhost:61005/v1"

Configuration Notes

  • If your model requires authentication, replace dummy_key with your actual API key
  • Replace localhost with the appropriate hostname if deploying on a remote server

Step 3: Test the Integration

Test your setup with the following command:

Test Command
uv run main.py common-benchmark --config_file_name=agent_llm_mirothinker output_dir="logs/test"

This command will:

  • Use the agent_llm_mirothinker configuration with the dedicated MiroThinkerSGLangClient
  • Run the example dataset benchmark (configured in the YAML file)
  • Test the model's question-answering capabilities

Configuration Details

The ./config/agent_llm_mirothinker.yaml configuration file uses:

  • provider_class: "MiroThinkerSGLangClient" - A dedicated client for MiroThinker models deployed with SGLang
  • Model path and generation parameters (temperature, top_p, max_tokens, etc.)
  • Environment variables for API endpoint configuration

Documentation Info

Last Updated: September 2025 · Doc Contributor: Xalp @ MiroMind AI