oQ beats mlx-lm on KL, RAM

oQ looks like the stronger default if KL divergence is your quality yardstick; it trades a modest RAM increase for a much cleaner approximation of the source model.

• –At 2-bit and 3-bit, oQ is dramatically better than mlx-lm’s Q in KL terms, which is where quantization usually hurts most.
• –By 6-bit and 8-bit, the gap narrows, so the decision becomes more about RAM budget than fidelity.
• –The MXFP4 and MXFP8 reference points are useful, but they do not change the basic story: sensitivity-aware allocation wins on distribution preservation.
• –The result reinforces the post’s broader point that “smallest file size” is not the same as “best quantization” for LLMs.