Benchmarking Local Qwen 3.6 Quantizations Against Frontier LLMs on a Complex Coding Primitive

This analysis investigates the capability of locally run, quantized models (Qwen 3.6 variants) when tasked with a highly complex, dense coding primitive: generating a single-file HTML canvas animation. The task required simulating a realistic, parallax-scrolling side-view of a moving car using only vanilla JavaScript and HTML, comparing the results against leading proprietary frontier models.

Experimental Setup and Task Definition

The experiment aimed to test the practical coding fluency of local Large Language Models (LLMs) against highly optimized, web-based frontier models. The chosen primitive was highly demanding, requiring not just basic code generation, but sophisticated animation logic, physics simulation (subtle chassis motion, spinning wheels), and complex graphical rendering (parallax scrolling, layered scenery, cinematic lighting).

The Prompt Specification

The generative prompt required the model to produce a complete, single-file HTML implementation. The specifications included:

• Full-page canvas implementation with zero external libraries.
• Simulation of a realistic side-view car.
• Continuous background scrolling and layered scenery (nearby ground, roadside elements, trees, distant hills) to create a natural parallax effect.
• Realistic wheel spinning and subtle body motion for immersion.
• Seamless looping animation with cohesive, cinematic lighting (sunset, dusk, or daylight).

Model Comparison and Performance Metrics

The test involved running the exact prompt across a set of both web-based frontier APIs and locally quantized models running on a specific hardware configuration (Ryzen 5 5600