Laya ships a 421M open decision model

Convai Innovations released Laya, a 421-million-parameter Apache 2.0 decision model that reports 32.8ms p50 latency on a Tesla T4, with a 322-million-parameter multilingual checkpoint for more than 100 languages. Its official release reports 0.766 accuracy on a typed-decisions benchmark, but says that result uses fine-tuning while its base models score near 0.35 zero-shot; the same page also recommends keeping choice schemas below 20 options. Builders routing tickets, guardrails, or triage can test this non-generative specialist locally, but should measure their own labels and acceptance rate rather than promote a benchmark win into a production guarantee; MLPerf’s RAG analysis separates benchmark qualification from application accuracy.