What is the best model you can actually run, for your use case, on your hardware,
under your rules? Ranked here in your browser — no account, no key, no server.
Each result names the evidence it rests on
(verified, partial-verified,
mixed, unverified-legacy, or
none) and the share of this use case's benchmark weight that
was actually measured. The wizard orders a model when measurements cover at least
25% of that weight and at least two benchmarks; thinner
cards stay visible as unranked. The CLI uses a 50% floor.
Treat the shortlist as something to investigate, not a verdict.