Julia 1 puts structured decisions on the CPU

Supersonic Labs released Julia 1, a 144.3-million-parameter Apache 2.0 model that chooses among supplied answers on a CPU; its own Banking77 pilot scored 64 of 100 after narrowing 72 labels, versus an 87% Jev reference. The lab’s release says the checkpoint uses 550.5 MiB and reached 73.15% on 2,000 typed decisions, but warns that long, similar option lists can discard correct answers. Builders can test it as a local router—not a chatbot—beside the archive’s cost analysis of typed decision models, keeping human review for consequential choices.