Q3 was confident and wrong

I quantized a 14B instruct model down to Q3 because the sample 24 GB card was already holding something else. The speed was lovely. The answers were not.

The fixture is twenty questions with an answer I can check without leaving the desk. Public dates, short arithmetic, and “which of these three is the odd one.” No private documents. Q5_K_M got 17. Q3 got 11, and the misses arrived in the same calm voice as the hits.

What fixed it was dull. I stopped sharing the card. Q5_K_M fit once the other model was unloaded, and the drop in tokens per second was smaller than the drop in correct answers had been. Q4_K_M was the compromise when I wanted both files loaded for a comparison.

Q3 stays on the shelf for “does this file boot.” It does not get the paragraph I might believe.

The numbers above are the sample fixture from this note, not a leaderboard. A different prompt would move them. The direction did not feel subtle: the smaller quant was sure, and it was worse.