Our summary
Some teams want AI decisions without sending their data to a hosted service. David Krcek tests Laya, an open local model, against earlier Jev results on banking questions. His comparison asks both how often the answer is right and what changes when the options or language change.
The test uses two hundred English questions across ten categories, then a hundred German translations. Laya also receives the English options in three orders. The author sends one unchanged Jev-shaped request to Laya's local server and gets matching answer fields, showing that existing calling code can be reused in that example.
In the author's tests, Laya answers faster locally but gets fewer labels right, especially in German, and changing option order changes some answers. Hosted Jev time includes a network trip; local time excludes loading the model. These version-specific results were not independently reproduced, and matching response fields does not establish equivalent quality.
Key takeaways
- Test local and hosted choices on the same questions, while keeping network time and model-loading time distinct.
- Try different option orders and the languages your users actually write. The interface alone cannot tell you whether decisions are dependable.
- Choose cutoffs on your own examples. The author's uncertain-answer handoff is a proposal based on a small test, not a production guarantee.