Local Qwen 3.6 vs frontier models on a coding primitive: single-file HTML canvas driving animation – results and GIFs

Saw this post comparing Qwen 3.6 variants on coding primitives, so I wanted to see how local quants stack up against frontier models on a similar dense, single-file coding task. I ran the exact same prompt across local and web-based models accessed through my Perplexity subscription.

The prompt

"Write a single HTML file with a full-page canvas and no libraries. Simulate a realistic side-view of a moving car as the main subject. Keep the car visible in the foreground while the background landscape scrolls continuously to create the feeling that the car is driving forward. Use layered scenery for depth: nearby ground, roadside elements, trees, poles, and distant hills or mountains should move at different speeds for a natural parallax effect. Animate the wheels spinning realistically and add subtle body motion so the car feels connected to the road. Let the environment pass smoothly behind it, with repeating but varied scenery that makes the movement feel believable. Use cinematic lighting and a cohesive sky, such as sunset, dusk, or daylight, to enhance atmosphere. The overall motion should feel calm, immersive, and realistic, with a seamless looping animation."

Models tested

Frontier (web-based via Perplexity, tok/s not measured):

• Claude sonnet 4.6 Thinking — used internet for reasoning
• Gemini 3.1 Pro Thinking
• GPT 5.4 Thinking
• Kimi k2.6 Thinking

Local (Ryzen 5 5600, 24 GB DDR4-3200, RX 5700 XT 8GB):

• Qwen3.5 9B Q4_K_M — ~50 tok/s
• Qwen3.6-27B (Claude-opus-reasoning-distilled) Q4_K_M — 2.65 tok/s
• Qwen3.6-27B Q4_K_M — 2.70 tok/s
• Qwen3.6-31B A3B Q4_K_M — 12.13 tok/s
• Gemma-4-31b-it — 1.91 tok/s
• Qwen3.5 4B Q8 — 60 tok/s — used internet for reasoning
• Qwen3.5 4B Q4_K_M — 80 tok/s — used internet for reasoning

What I looked for
Realistic side-view driving animation: layered parallax scenery, spinning wheels, subtle chassis motion, cohesive sky and lighting, and seamless looping — all vanilla JS/canvas, zero libraries.

Subjective ranking for this specific task

1. Kimi k2.6 Thinking — cleanest overall visual result
2. Qwen3.6-27B Q4_K_M (local) — stronger than I expected; good parallax and road feel
3. Qwen3.6-27B Claude-opus-reasoning-distilled — close third

The local 27B quant delivered more natural motion and layering than some frontier outputs for this specific visual primitive. I was expecting frontier models to do much better — am I missing something?

Outputs
I only changed the HTML <title> tags to track which model generated which file. I’ll share all the output files and probably a few screenshots of the running animations so you can judge the visual quality yourself.

If anyone wants to run the exact same prompt on their setup — especially other MoE cuts or distills — feel free to share your results.