Super interesting test by @atomic_chat_hq

Shows that more quantization precision does not automatically produce the better result.

Here, Qwen 3.8 27B Q8 (8-bit quantized version of the model) sometimes matched or beat Q4 (4-bit quantized) on the same voxel tasks.

They finally recommend AD-Q5_K_M as their preferred quant, since it runs on a 32GB MacBook Air with 32K context and retains 97.3% next-token agreement with BF16.

For the model weights alone:

Q8_0: 28.9GB
AD-Q4_K_M: 17.1GB

So Q4 saves 11.8GB of RAM/VRAM, or about 40.8% versus Q8.