Qwen 3.5 35B vs Sonnet 4.5: Is The Gap CLOSING?

The video introduces the new Qwen 3.5 Medium Series from Alibaba, which features a 35B parameter model designed to run on consumer hardware like a MacBook Pro. These models claim to beat Anthropic's Sonnet 4.5 in benchmarks, a claim the narrator aims to verify through hands-on testing. While Sonnet 4.5 is praised for its premium integration with the Anthropic suite, it is noted as being expensive and requiring an online connection. Qwen 3.5 35B represents a shift toward powerful local LLMs that could potentially democratize high-level coding assistance. The creator sets the stage for a three-part testing gauntlet covering easy, medium, and hard coding challenges.

The first test involves building a basic to-do list application using React and Vite. Sonnet 4.5 produced a functional, aesthetically pleasing app with local storage persistence, though its code structure was slightly cluttered. Qwen 3.5 initially appeared to win by including advanced features like categories and severity levels with a cleaner component architecture. However, the narrator discovered that Qwen had an unfair advantage through enabled 'superpower skills' in the OpenCode environment. When re-tested without these aids, Sonnet 4.5 proved to be the more reliable and consistent model for the task.

The medium-difficulty task required the models to create a 3D interactive solar system using 3JS. Sonnet 4.5 successfully generated a functional scene with smooth rotation, panning, and interactive planetary data in a single attempt. In contrast, Qwen 3.5 failed significantly, producing a blank page and numerous console errors that it could not resolve even after multiple prompts. The model also displayed erratic behavior, such as 'falling asleep' and creating redundant, messy file structures with incorrect directory nesting. This section highlights a significant gap in the models' ability to handle complex libraries and multi-step logic.

For the final hard test, both models were tasked with modifying an existing codebase to add a tweet-to-screenshot capture feature. Sonnet 4.5 initially hit a timeout error but was able to fix its own bug and successfully deliver the image download functionality. Qwen 3.5 attempted the task but failed to overcome a 60-second timeout issue, even after being explicitly asked to extend the limit. It struggled to identify the root cause of the execution failure, whereas Sonnet demonstrated better self-correction and debugging capabilities. This confirms that Sonnet 4.5 remains the superior choice for enterprise-level or complex development workflows.

The creator concludes that while Qwen 3.5 is impressive for a 35B parameter local model, it cannot yet compete with the likes of Sonnet 4.5. He theorizes that Qwen's high benchmark scores are likely the result of intensive training on specific evaluation datasets rather than general reasoning improvements. The discussion points out that Sonnet likely utilizes a much higher parameter count and more diverse training data, giving it a 'reasoning' edge that benchmarks miss. This section emphasizes the importance of performing real-world evaluations over relying solely on marketing numbers. Finally, the video mentions Qwen's separate successes in the text-to-speech (TTS) space as a secondary area of interest.