Which local model for your GPU?
Almanac is the open-source local AI companion for Final Fantasy XIV (github.com/Spaceghost/almanac-dalamud), made by Johnneylee Jack Rollins. It runs a model on your own machine. These results come from players who ran its benchmark suite and chose to share them, so you can pick the model that gives the best answers for the memory and speed you have.
Best models by VRAM tier
Grouped by the dedicated VRAM of the card the run was measured on. Click a column heading to sort. Score is the suite's weighted score (0-100) with its 95% range; who counts distinct submitters.
Loading…
Quality vs speed
One dot per model and quantisation in the chosen tiers: higher is better answers, further right is faster generation (log scale). Hover or focus a dot for details; the tables above hold the same numbers.
By GPU family
Privacy. Sharing is opt-in and needs a sign-in with GitHub or FFXIV (XIVAuth): Almanac (0.2.0 or newer, plugin and command line) shows a short code and you approve it on the connected apps page. With each result the server keeps which sign-in you used, a key derived from that account's id and which linked app sent it; none of that is shown on the leaderboard, and no login name, e-mail or provider token is kept. A result holds the suite, the model, the backend, the GPU's marketing name, its VRAM, the operating system family and the scores; the server refuses anything else, and refuses text that looks like a file path, an address or an e-mail. Your IP address is not stored: a keyed digest of it counts submissions for the rate limit, and each account counts once per model. The site owner can hide results that look wrong. To have a result removed, contact the owner through github.com/Spaceghost.
This page is counted, without cookies and without anything that identifies you: what this site counts.