This video provides an in-depth review and practical testing of the newly updated Alibaba Cloud Quen 3-235B-822B-instruct AI model, designated "257." The updated version features a significantly expanded context length and a new "non-thinking" mode that delivers direct outputs without internal processing thoughts. The reviewer tests the model's capabilities through various coding challenges, including generating a browser-based operating system and a 3D racing game, which yield mixed results initially but improve with iteration. The model demonstrates exceptional long-context comprehension with "The Great Gatsby" and produces a remarkably creative and insightful abstract comparison between the novel and the consumer GPU market. While a complex digital audio workstation proved challenging, a simpler drum sequencer and keyboard-controlled piano showed functional success. The reviewer concludes that the "257" variant is a highly interesting and impressive model, especially for creative and conceptual tasks, benefiting from its increased context and streamlined output.
Exploring the Updated Alibaba Cloud Quen AI Model (257 Variant)
0:16 Introducing Quen 3-235B-822B-instruct (257): The video introduces Alibaba Cloud's newly updated AI model, version "257," highlighting its potential based on benchmark JPEGs and the intent to test it in its most powerful configuration via the official chat.quen.ai interface.
1:20 Significant Context Length Improvement: A major technical enhancement is the native context length, which has been dramatically increased from 32,768 to 262,144 tokens, a very substantial improvement for handling large amounts of information.
2:22 Non-Thinking Mode: The "257" variant supports only "non-thinking" output, meaning it does not generate internal thought processes or "think blocks." This provides immediate answers, which the reviewer finds preferable for directness.
2:55 Enhanced General Capabilities: The model boasts significant improvements in general capabilities, longtail knowledge coverage across multiple languages (supporting 119 known dialects), and better alignment with user preferences for subjective and open-ended tasks.
4:04 Benchmark Honesty: The reviewer notes an interesting detail: the model's benchmark results accurately show a slight performance dip in the "Ader Polyglot" coding benchmark compared to its previous iteration, indicating transparent reporting.
4:30 Browser-Based OS Generation (Initial Attempt): The first coding task involved creating a browser-based operating system using HTML, JavaScript, and CSS. The initial 930-line output had visual flaws (blank white background) and non-functional applications (e.g., notepad couldn't type, calculator buttons were unresponsive).
7:20 Browser-Based OS Generation (Second Attempt): After a re-prompt, the model successfully fixed the background with an aesthetic image and made the notepad and calculator applications functional (though calculator buttons had poor contrast and windows were not resizable by dragging). A functional web browser application was also included.
10:39 Low-Poly 3D Racing Game Generation (Mixed Results): The model was tasked with generating a low-poly 3D racing game. The first attempt had visual issues (menu remained visible, distorted view, odd opponent graphics) but showed functional speed, position, and clock. The second attempt to fix these issues resulted in a "significantly worse" output.
12:51 "The Great Gatsby" Context Length Test: To evaluate the extended context length, the entire text of "The Great Gatsby" was input. The model successfully provided a summary and accurately answered specific questions about the plot, demonstrating strong long-context comprehension.
14:06 Abstract Comparison Task (Creative Success): A challenging creative prompt asked the model to relate "The Great Gatsby" to the consumer GPU market at a "PhD level." Initially, the model "refused" in five paragraphs, citing the irrelevance. However, when re-prompted as a "test of creativity" for a "poetry PhD assignment," it produced an "extremely good" and insightful five-paragraph comparison, drawing profound analogies (e.g., Gatsby as Nvidia, Valley of Ashes as semiconductor fabs).
17:26 Character-Company Analogies: As part of the creative test, the model was asked to name a computer company analogous to each main character: Jay Gatsby as Nvidia, Tom Buchanan as IBM, Daisy Buchanan as Apple, and Nick Carraway as AMD.
18:05 Web-Based Digital Audio Workstation (DAW) and Drum Sequencer: An attempt to create a web-based DAW yielded a visually impressive but non-functional result. Simplifying the prompt to a drum sequencer was successful, producing a functional and aesthetically pleasing interface.
20:29 Keyboard-Controlled Piano Integration: A final addition of a keyboard-controlled piano to the drum sequencer was mostly successful, offering different wave types with unique sounds, though key mapping was off and the kick drum in the sequencer ceased to function.
21:56 Overall Conclusion: The reviewer concludes that the "257" variant is a "very interesting and impressive model," particularly highlighting its performance in the obscure "Great Gatsby" comparison and the practical benefits of its extended context and direct "no thinking" output.