Julia 1 matches Jev on 73.15% benchmark, loses 87% vs 61%

The comparison matters because Jev and Julia 1 approach the same "judgment model" task from opposite directions. Jev, from TypeSafe AI, is a cloud API opened to early access on September 15, 2026; Julia 1, from Brazil's Supersonic Labs, was published on Hugging Face on September 25-26, 2026, with weights under Apache-2.0. Both take a state, a question and options, and return a probability per option, but Jev is only reachable through a paid cloud API while Julia 1 runs on a CPU.

The reproduction itself carries weight. The author re-ran the 2,000-question typed-decisions test on a MacBook Pro with an Apple M5 Max and got 72.55%, matching the official CPU re-measurement exactly; the gap from Jev's 72.70% is within noise. That said, Julia 1's README does not say whether its training overlapped with the benchmark tasks, and Jev's figures were transcribed from third-party records measured on different days rather than a head-to-head run.

The author also reports that Julia 1's training and experiments used about $104 (540 reais) in cloud GPU costs, and that Julia 2 is planned to be built from the ground up rather than on mmBERT-small. The outcome likely hinges on whether Julia 1's accuracy holds on each user's own data once the question wording and option descriptions are adjusted, which is why the author's own recommendation is to measure 50-100 real inquiries before choosing.