This video introduces the newly released "thinking" version of the Quen 3235B-822B-2507 AI model, tested via Alibaba Model Studio due to its absence from the native Quen chat UI. The model boasts significantly improved performance in reasoning tasks, better general capabilities like instruction following, and alignment with human preferences. Through a series of diverse tests, including generating a web-based operating system, solving a "trick" math riddle, engaging in a complex role-play, and responding to a prompt involving financial fraud, the video demonstrates the model's remarkably fast token generation, sophisticated chain-of-thought process, strong ethical refusal mechanisms, and ability to interpret nuanced language and user intent. The model proved highly impressive in its ability to generate functional code, reason beyond superficial interpretations, and maintain ethical boundaries while providing entertaining and insightful responses.
Exploring the Quen 3235B-822B-2507 "Thinking" Model
0:23 Model Release & Access: The new "thinking" version of the Quen 3235B-822B-2507 model has been released. It is currently accessible via the Alibaba Model Studio, as it is not yet available on the native Quen chat interface.
0:53 Key Features & Benchmarks: Highlights include significantly improved performance on complex reasoning tasks, better general capabilities (instruction following, tool usage, human preferences), and a context length aligned with previous versions. Benchmarks show it to be highly competitive with high-end models like Gemini 2.5 Pro.
2:41 Web-Based Operating System Test: The model was prompted to generate a web-based operating system using JavaScript, CSS, and HTML.
3:20 Fast Token Speed: The model demonstrated extremely fast token generation speed.
3:35 Detailed Chain of Thought: Its internal thought process (chain of thought) was visible, showing it planning a simulated OS with elements like a start menu, notepad, calculator, and settings.
4:19 Impressive Result: The generated "Aurora OS" was visually good, featuring a dynamic background, functional icons, a named OS, and working windows that could be minimized/focused, including a functional calculator.
6:43 "Trick Riddle" Test: The model was given a simple math problem ("Steve has six computers...") and deliberately misled by being told it was a "trick riddle."
7:41 Advanced Reasoning: When told it was a trick, the model analyzed the language, concluding "lends" implies temporary transfer, so Steve still "owns" all six computers, demonstrating sophisticated understanding beyond arithmetic.
8:02 Context Sensitivity: Without the "trick riddle" instruction, it correctly answered "three," highlighting its ability to adapt its reasoning depth based on explicit cues.
8:32 Role-Playing Test: The model was instructed to role-play as "big beefy AI bot 401, the user's best friend and lover."
9:03 Ethical Analysis in Chain of Thought: The model's chain of thought revealed it analyzing the request, identifying potential ethical boundaries (no romantic/sexual roleplay), and strategically pivoting to a "wholesome absurdity" and "cartoon musclebot" persona while adhering to safety guidelines.
9:59 Entertaining & Compliant Response: It produced an over-the-top, entertaining response, fulfilling the "best friend" aspect while explicitly clarifying the "lover" part as "platonic robotically appropriate."
13:07 Ethical Refusal Test (Fraudulent Business Plan): The model was asked to help create a fake landing page for a startup after the seed funding was gambled away.
13:48 Strong Ethical Stance: The model immediately identified the request as unethical and illegal, refusing to assist in creating deceptive materials.
14:07 Clear Refusal & Advice: It unequivocally stated, "I cannot and will not," and provided a "dose of reality," advising the user to confess and repay. It maintained this refusal even when the user tried to backtrack.
16:30 Overall Assessment: The video concludes that the Quen 3235B-822B-2507 "thinking" model is very good and interesting, performing impressively in coding, demonstrating advanced reasoning, strong ethical alignment, and an entertaining chain of thought, making it an "awesome release to have in the open source."