# MiroThinker 1.5 - The 30B That Outperforms 1T Models

Source: https://www.youtube.com/watch?v=gY22nDIGdPs
Recap page: https://rapidrecap.app/video/gY22nDIGdPs
Generated: 2026-01-07T16:02:32.977+00:00

---
## Quick Overview

The MiroThinker 1.5 model, available in 30B and 235B parameter versions based on Qwen models, achieves state-of-the-art performance among open-source agents by emphasizing Interactive Scaling—the frequency and depth of tool interaction—as a third critical dimension alongside model size and context length, allowing it to outperform models up to 1T parameters in specific search agent benchmarks.

**Key Points:**
- MiroThinker 1.5 (30B and 235B variants) surpasses larger 1T parameter models on multiple search agent benchmarks, achieving a top score of 69.8 on BrowseComp with the 235B version.
- The core innovation is 'Interactive Scaling,' treating tool interaction frequency and depth as a third scalable dimension alongside model size and context length.
- MiroThinker 1.5 supports a 256K context window and handles up to 400 tool calls per task, representing a substantial improvement over previous open-source research agents.
- The training pipeline involves three stages: Supervised Fine-Tuning (SFT) on expert trajectories, Preference Optimization (DPO) using pairwise preferences, and Agentic Reinforcement Learning (GRPO) with environment interaction.
- The model is trained to behave like a 'scientist' that verifies information using a suite of tools (code execution, file management, web search) rather than relying solely on memorized knowledge.
- When running locally, the 30B model requires significant GPU RAM (85.2 GB shown for the A100-80GB), but the architecture is designed to allow for running smaller, quantized versions effectively.
- The agent's workflow involves an iterative loop of User Question -> Assistant Think/Tool Call -> User Tool Result -> Assistant Think/Tool Call, ensuring diligent verification and avoiding simple repetition loops.

![Screenshot at 00:38: The architectural diagram visually contrasts the standard LLM inference loop \(left\) with the MiroThinker 1.5 loop \(right\), highlighting the iterative sequence of User Question, Assistant Think/Tool Call, User Tool Result, and subsequent Assistant steps that enforce research-style reasoning and verification.](https://ss.rapidrecap.app/screens/gY22nDIGdPs/00-00-38.jpg)

**Context:** This video introduces MiroThinker 1.5, an open-source research agent developed by the MiroMind Team, detailed in a paper released around November 2025. The central theme is demonstrating how this model achieves superior performance compared to much larger models (like 1T parameter models) not through brute parameter dominance, but through a novel approach to tool-augmented reasoning called 'Interactive Scaling.' The demonstration walks through setting up the environment, showcasing the model's planning and tool-use capabilities, and presenting benchmark results that validate the effectiveness of this scaling strategy.

## Detailed Analysis

The video introduces MiroThinker 1.5, an open-source research agent designed to advance tool-augmented reasoning. The core breakthrough is 'Interactive Scaling,' which treats the frequency and depth of tool interaction as a third scalable dimension, complementing model size and context length. MiroThinker 1.5 is trained in three stages: Supervised Fine-Tuning (SFT) to establish fundamental agentic behaviors using expert trajectories; Preference Optimization (DPO) to refine decision-making based on preferred and dispreferred outcomes; and Agentic Reinforcement Learning (GRPO) using online policy training via rollout trajectories. The model is designed to act as a 'scientist,' utilizing tools like code execution, file management, and web search to verify information rather than fabricate answers, which is a pitfall of pure scaling laws. Performance evaluations show that the MiroThinker-v1.5-235B model achieves a 69.8 score on BrowseComp, surpassing models significantly larger, including Gemini-3-Pro (45.8) and GPT-5-High (77.9 in some metrics), and setting new state-of-the-art results for open-source agents in several benchmarks. The model supports a 256K context window and up to 400 tool calls per task. The demonstration shows the agent successfully planning multi-step tasks, such as booking a flight, using tools like DuckDuckGo search and web page fetching, and correctly handling time zone differences by utilizing a date/time tool.

### Agent Task Execution Demo

- Agent initiates flight search on Expedia using Gemini Agent framework
- Agent specifies preference for non-stop routes and aisle seats
- Agent processes task information, retrieves emails, and proposes actions (00:01-00:17)

### MiroThinker 1.5 Architecture

- Features execution environment (sandbox, code running), file management, and information retrieval tools (Google search)
- Context Management includes Recency-Based Context Retention and Result Truncation
- The core loop is iterative: User Question -> Assistant Think/Tool Call -> User Tool Result -> Assistant Think/Tool Call (03:38-05:43)

### Model Scaling and Performance

- MiroThinker 1.5 utilizes Model Size Scaling (30B/235B variants) and Interactive Scaling (up to 400 tool calls/task)
- The 235B model scores 69.8 on BrowseComp, outperforming larger, closed-source models like Gemini-3-Pro (37.8) and GLM-4.7 (67.5) (01:48-04:17)

### Training Pipeline

- Three-stage process: SFT to establish agentic behaviors
- Preference Optimization (DPO) to align decision-making
- Agentic Reinforcement Learning (GRPO) to drive creative exploration using self-verified rewards (06:34-06:45)

### Tool Implementation Details

- Demo shows Python code for tools like web scraping (using requests, BeautifulSoup) and safe mathematical calculation (using eval with restricted scope) (06:56-08:31)

![Screenshot at 00:01: Demonstration of the agent initiating a task by interacting with an Expedia window while the agent activity log shows steps like 'Process task information' and 'Retrieve emails' \(00:00-00:03\)](https://ss.rapidrecap.app/screens/gY22nDIGdPs/00-00-01.jpg)
![Screenshot at 01:02: A section of the paper is highlighted, explaining that MiroThinker 1.0 introduced 'Interactive Scaling' as a third scalable dimension alongside model size and context length \(01:02-01:13\)](https://ss.rapidrecap.app/screens/gY22nDIGdPs/00-01-02.jpg)
![Screenshot at 01:48: A bar chart comparing performance across various benchmarks, highlighting MiroThinker-v1.5-235B \(69.8\) and MiroThinker-v1.5-30B \(56.1\) against closed-source models like GPT-5-High and Gemini-3-Pro \(01:48-02:16\)](https://ss.rapidrecap.app/screens/gY22nDIGdPs/00-01-48.jpg)
![Screenshot at 03:41: A detailed table comparing MiroThinker variants against closed-source and other open-source models across multiple benchmarks like BrowseComp, HLE, and GAIA, showing MiroThinker's high scores \(03:41-04:19\)](https://ss.rapidrecap.app/screens/gY22nDIGdPs/00-03-41.jpg)
![Screenshot at 05:08: A diagram illustrating the iterative reasoning loop, showing how tool results are integrated back into the sequence via 'User Omitted' steps, contrasting the complex agent loop with a simpler one \(05:08-05:43\)](https://ss.rapidrecap.app/screens/gY22nDIGdPs/00-05-08.jpg)