mlx-kld benchmarks oQ, Q, MXFP, UD quants

Hot take: this is the right way to evaluate MLX quantization. Once you look at divergence from the reference distribution instead of just memory savings, it becomes obvious that “4-bit vs 6-bit” is too crude a shortcut.

• –KLD is a cleaner signal than perplexity for quantization damage because it isolates the effect of the quantizer itself.
• –The useful takeaway is not a single winner, but that oQ, native Q, MXFP, and UD can trade places depending on the model and architecture.
• –For dense models, higher-bit or better-targeted 6-bit schemes look like the safest default; for MoE models, router-sensitive tensors can dominate the outcome.
• –The benchmark reinforces that MLX users should treat quantization as a per-model decision, not a one-size-fits-all setting.
• –The project is also practical: caching reference log-probs makes this kind of comparison feasible on Apple Silicon instead of purely theoretical.