An open-weight listing is a promise you can download something, not a promise you can run it. The three listings below are here because they span the range a self-hoster actually faces: a weight size that fits one card, a size that needs care, and a size that needs hardware you probably do not have. Read them as a spectrum rather than a ranking.
The four checks, in order
License first. If the license forbids the way you intend to serve it, nothing else on the page matters. Second, size class: parameter count is not a quality score, it is a memory estimate. Third, context window, because the window you need decides the memory you need on top of the weights. Fourth, the VRAM figure where a source provides one, treated as a floor, not a guarantee.
What the highlighted cells do not mean
The comparison table marks the best number in each row. A model can win every numeric row and still be the wrong choice, because the rows do not include the two things that usually decide it: whether the vendor answers a support ticket, and whether the license matches your business. Read the table as a filter, then read the model card.
After you pick
Pin the weights and record the file hash next to the license note. A later download with a different hash is a different model even if the card title did not move. Re-run the one prompt you graded before you let the new file replace the old build, and record the result. The open source and proprietary models guide covers the five lines to hand a teammate when you make this call.