Compare text models

Pick up to three runs and see them side by side: speed, quality with its margin of error, memory, the world each one built, and who made the recipe. Every number is the one on the run's own page, and it links to its raw output.

Mia Flash-Next 1 GB10 box, ThinkingCap 27B 1 GB10 box, Mia DeepSeek 2 GB10 boxes →

Overall scores are relative to the best run of each memory class on the date shown, so two classes are not on one scale: compare them by the raw numbers. A cut-off answer counts as wrong. Thinking is off except in the last quality test. Specialist models are marked; my tests do not measure that strength. A run only has an overall score and a browser check after I rated the page it built.