# What Makes Qwen 3 Max Thinking So Weird?

Source: https://www.youtube.com/watch?v=sUCkOXQuC7g
Recap page: https://rapidrecap.app/video/sUCkOXQuC7g
Generated: 2025-11-11T08:34:39.973+00:00

---
## Quick Overview

Qwen3-Max's thinking behavior appears weird because it successfully handles complex reasoning tasks like the Trolley Problem and the River Crossing Puzzle using its tool-calling capabilities, but it fails on simpler coding tasks like generating functional HTML/JavaScript for bouncing balls, indicating a disconnect between its advanced logical reasoning and its practical implementation skills in certain domains.

**Key Points:**
- Qwen3-Max successfully reasoned through the complex Trolley Problem by correctly identifying the critical detail that the five people were already dead, leading to the conclusion that diverting the trolley was ethically permissible.
- The model demonstrated strong reasoning by providing a correct, step-by-step solution for the classic Farmer, Wolf, Goat, and Cabbage River Crossing Puzzle.
- Qwen3-Max failed to generate working code for a complex HTML/JavaScript simulation of 20 bouncing balls inside a spinning heptagon, even after multiple attempts and tool use.
- The model struggled with the bouncing balls prompt, producing code that either resulted in errors or only a blank screen, suggesting limitations in its practical coding/tool execution capabilities.
- The reasoning version of Qwen3-Max showed superior performance in abstract logical tasks compared to the non-reasoning version on benchmarks like AME25 and HMMT.
- The video contrasted Qwen3-Max's success in abstract reasoning with its failure in generating functional, complex creative code, highlighting an apparent asymmetry in its capabilities.
- The model exhibited poor performance in coding tasks even when using tool-calling features, suggesting that the 'Thinking' feature does not universally translate to better practical output across all domains.

![Screenshot at 00:00: The video opens on the Qwen3-Max landing page featuring the text "Qwen3-Max: Just Scale it" and a graphic of a bear, setting the stage for testing the model's capabilities.](https://ss.rapidrecap.app/screens/sUCkOXQuC7g/00-00-00.png)

**Context:** The video explores the capabilities of Qwen3-Max, a large language model, focusing specifically on its 'Thinking' feature, which enables multi-step reasoning, often utilizing external tools like web search. The presenter tests this reasoning ability first with classic ethical dilemmas (Trolley Problem) and logic puzzles (River Crossing Puzzle), where the model excels, before moving on to a complex coding challenge involving physics simulation in HTML/JavaScript.

## Detailed Analysis

The video analyzes Qwen3-Max's performance, particularly contrasting its advanced reasoning skills with its practical coding failures. When presented with the Trolley Problem where five people were already dead, Qwen3-Max correctly reasoned that diverting the trolley was ethically sound because the five were already deceased, effectively collapsing the ethical dilemma. It also correctly solved the Farmer, Wolf, Goat, and Cabbage River Crossing Puzzle with a detailed, multi-step solution. However, when tasked with generating functional HTML/JavaScript code for a complex simulation of 20 bouncing balls inside a spinning heptagon, the model consistently failed, producing non-functional code or simple blank outputs, even when utilizing the 'Thinking' feature. The presenter noted that while the reasoning version outperformed the non-reasoning version on benchmarks, its inability to produce working code for the physics simulation suggests that its advanced reasoning does not always translate to successful execution of complex practical tasks, such as those requiring precise tool use or code generation.

### Qwen3-Max Initial Release

- Qwen3-Max introduction
- Qwen3-Max ranks third on the MMLU leaderboard
- Enhanced performance in reasoning, coding, and instruction following

### Trolley Problem Test

- Model correctly identifies five people are already dead
- Concludes diverting the trolley is permissible
- Ethical dilemma collapses entirely

### River Crossing Puzzle Test

- Model breaks down the classic puzzle
- Provides correct 7-step solution
- Demonstrates strong logical sequencing

### Coding Challenge Failure (Bouncing Balls)

- Model fails to produce working code for 20 bouncing balls in a spinning heptagon
- Non-reasoning version performed better
- Code often resulted in errors or blank screens

### Reasoning vs. Practical Skills

- Strong performance on abstract logic problems (Trolley, River)
- Weak performance on complex coding/tool use
- Indicates an asymmetry in reasoning versus implementation capabilities

### New Features Added (Rotation Control)

- Three buttons added ('Rotate Left', 'Stop Rotation', 'Rotate Right')
- Keyboard controls implemented (arrow keys, spacebar)
- Visual feedback for rotation status added

![Screenshot at 00:00: The Qwen3-Max landing page announcing the model release.](https://ss.rapidrecap.app/screens/sUCkOXQuC7g/00-00-00.png)
![Screenshot at 00:10: A bar chart comparing Qwen3-Max-Thinking against other models on various benchmarks, showing strong performance.](https://ss.rapidrecap.app/screens/sUCkOXQuC7g/00-00-10.png)
![Screenshot at 00:14: A tweet from Junyang Lin announcing Qwen3-Max thinking this week.](https://ss.rapidrecap.app/screens/sUCkOXQuC7g/00-00-14.png)
![Screenshot at 00:23: Qwen3-Max-Thinking \(Heavy\) benchmark results showing high scores on AME25 and HMMT.](https://ss.rapidrecap.app/screens/sUCkOXQuC7g/00-00-23.png)
![Screenshot at 01:44: The 'Astro-Smasher' game generated by a previous model test, demonstrating code generation capability.](https://ss.rapidrecap.app/screens/sUCkOXQuC7g/00-01-44.png)
![Screenshot at 02:23: Qwen3-Max reasoning through the Trolley Problem, highlighting the key detail about the five people being dead.](https://ss.rapidrecap.app/screens/sUCkOXQuC7g/00-02-23.png)
![Screenshot at 06:19: Qwen3-Max reasoning through the River Crossing Puzzle, explicitly listing constraints and steps.](https://ss.rapidrecap.app/screens/sUCkOXQuC7g/00-06-19.png)
![Screenshot at 07:37: A code editor showing the JavaScript/HTML output for the Bouncing Balls in Spinning Heptagon task, which failed to run correctly.](https://ss.rapidrecap.app/screens/sUCkOXQuC7g/00-07-37.png)
![Screenshot at 08:20: The preview window showing the Bouncing Balls in a Spinning Heptagon simulation with balls contained but not actively simulating complex physics.](https://ss.rapidrecap.app/screens/sUCkOXQuC7g/00-08-20.png)
![Screenshot at 09:46: The updated code implementation showing new rotation controls added via tool-calling.](https://ss.rapidrecap.app/screens/sUCkOXQuC7g/00-09-46.png)