THX-01

322M multilingual encoder decision model that answers typed questions and returns supporting evidence: alongside choice / yes-no / ordinal it outputs numeric extractions, verbatim excerpts, and citations — the output contract of an extraction model in an encoder-sized decision package. Post-training covers 18 languages; all comparisons are producer-run, and the 1,024-token input limit (state plus question plus options) bounds it to short-context work. Apache 2.0.