Quantization split performance

On an M3 Max 128GB, 9 local models based on 8-bit MLX were compared using the same flight combat sim prompt.

All conditions were identical. Each was tasked with building a single-file HTML, offering a choice of jet/prop/wildcard 3 types, and implementing dynamic enemies, tracers, damage, and crash spiral. Results were evaluated by prompts-to-final and whether the game was actually playable.

There were three key findings.

• Provider differences mattered more than the number of quantization bits. Three 8-bit variants of the same Qwen3.6 35B produced different game quality.
• Unsloth finished fastest with 3 prompts and 1,304 lines, properly including a minimap and a round planet.
• MLX Community was decent with 4 prompts, while oMLX only barely got it right after debugging 5 times a problem where controls kept reverting to neutral.

Lines of code had almost no correlation with quality.

• The winner was Qwopus 3.5 27B, finishing in 2 prompts and 1,049 lines.
• In contrast, Qwen Coder Next 80B produced the most code with 3 prompts and 1,635 lines, but yielded worse results with an oversensitive camera, missing enemies, and a plane rotated 180 degrees.

Surprisingly, only Qwopus added actual flight physics that hadn't been requested, implementing thrust/drag, per-aircraft aerodynamic constants, and frame-based velocity damping.