While OpenAI’s Operator and Deep Research are still in limited beta, the open-source community isn’t sitting still. Recently, the MiroMind AI team released an open-source deep research Agent—MiroThinker 1.5, claiming to solve problems through reflection rather than just piling on parameters.

Unlike past models that improve performance simply by increasing parameter count, MiroThinker 1.5 introduces the concept of Interactive Scaling. This allows the agent to improve its ability to solve complex problems through repeated interaction with the environment, trial-and-error, and reflection.

MiroThinker 1.5 Technical Breakthroughs

MiroThinker 1.5 was released in 30B and 235B parameter sizes. Its core logic lies in extending intelligence from within the model to the external world, handling long-horizon tasks through a "Reasoning-Verification-Correction" loop.

1. Interactive Scaling Paradigm

Most current LLMs operate in "exam-taker mode," providing a one-shot answer. MiroThinker relies not only on the model’s internal knowledge but emphasizes multi-turn deep interactions within the environment.

• Long-horizon reasoning: Supports a 256K context window, enabling it to process massive amounts of information.
• High-frequency tool usage: Supports up to 400 tool calls per single task, far exceeding similar open-source agents.

2. Performance

According to official data, the MiroThinker-v1.5-30B version beats Kimi-K2-Thinking on the Chinese web comprehension benchmark (BrowseComp-ZH) with a smaller parameter size and just 1/20th of the inference cost.

The larger 235B version scored an impressive 80.8% on GAIA-Val-165 (a General AI Assistant benchmark), placing it in the top tier for benchmarks like HLE-Text and BrowseComp.

3. Time-Sensitive Training

To solve "spoiler hallucinations" in prediction tasks, the development team strictly followed causality during training, ensuring the model "can only see the past, not the future." This design gives MiroThinker real-world value in scenarios like financial prediction and market trend analysis. This is why it was able to successfully predict A-share limit-ups and GTA 6 release trends.

Fast Installation Guide

MiroThinker is a tool-heavy Agent, and the official documentation strictly requires Python 3.10+. To save you the trouble of configuring environment variables and to avoid crashing your local system due to version conflicts, I recommend using ServBay to manage your Python environment. This is the fastest way to get up and running.

Step 1: Prepare Python Environment

• Open ServBay and find Python in the "Packages" panel.

• Select Python 3.10 (or a higher version) and click the green button to install.

• Once installed, the Python environment is ready immediately—no need to manually tweak Path variables.

Step 2: Clone Code & Install Dependencies

With the environment ready, you can pull the code and install dependencies via the terminal:

# Clone the repository
git clone https://github.com/MiroMindAI/MiroThinker
cd MiroThinker

# Enter the Agent directory
cd apps/miroflow-agent

# Install dependencies (ensure uv is installed)
uv sync

Step 3: Configure API Keys

MiroThinker’s power relies on external tools (like search and code execution). You need to copy the configuration file and fill in the corresponding Keys:

cp .env.example .env

In the .env file, you need to configure at least these basic services:

• SERPER_API_KEY: For Google Search.
• JINA_API_KEY: For web content scraping and summarization.
• E2B_API_KEY: Provides a secure sandbox environment to execute Python code.
• SUMMARY_LLM: An LLM used for information extraction (you can use lightweight models like Qwen3-14B).

Start & Run

Once configured, you can run a simple task via the command line to test if the Agent is working properly:

# Run the main program using uv
# Assuming you have deployed the MiroThinker model locally or remotely (e.g., via vLLM)
uv run python main.py llm=qwen-3 agent=mirothinker_v1.5_keep5_max200 llm.base_url=http://localhost:61002/v1

If your local VRAM cannot handle the 30B model, you can use API mode (connecting to Claude or GPT) to experience its thought process first.

# Use Claude (requires configuring ANTHROPIC_API_KEY)
uv run python main.py llm=claude-3-7 agent=single_agent_keep5

Final Thoughts

MiroThinker 1.5 gives me the impression that the open-source community is finally starting to focus on "System 2" (slow thinking). Although configuring Keys is a bit tedious, watching the Agent step through "Reasoning-Verification-Correction" in the terminal feels very different from LLMs that just hallucinate answers.

If you have complex research tasks or want to study how to reduce AI hallucinations, use ServBay to spin up an environment in minutes and give it a try. After all, an AI that can "admit it doesn’t know and go verify" is what we really need.