What happened
| Quant | GPU residency | Approx speed |
|---|---|---|
| Unsloth UD-IQ4_XS | 100% GPU · ~15.3GB | ~43 coding tok/s |
| Library Q4 | spilled | ~18 tok/s |
| Video-like UD-Q4_K_M (~46 GPU layers) | spilled worse | ~7 tok/s |
Takeaway
Copying a YouTube layer count is not a plan on 16GB. Promote IQ4_XS for experiments; skip the 46-layer UD-Q4_K_M recipe on this card.
Sister site: toklanes.com — measured tok/s review-bench.