Convai Innovations releases open-source multilingual decision model
Laya is designed to make structured decisions across more than 100 languages in a single forward pass, with reported GPU latency of about 33 milliseconds.
Convai Innovations has released Laya, an open-source, non-autoregressive decision model designed to produce structured probability predictions and schema choices without generating text.
The company says Laya supports more than 100 languages across 22 writing systems and delivers predictions in a single forward pass. It reports GPU latency of about 33 milliseconds, falling to 7.2 milliseconds per question in batched use.
Laya is intended for structured tasks including classification, routing, guardrails and triage. Convai Innovations says the model includes language routing based on writing systems and Latin-language signals, with three specialised checkpoints consolidated in a single Hugging Face repository.
The model’s weights are distributed under the Apache 2.0 licence, and the release claims there are no API subscription costs. Convai Innovations says Laya builds on research, datasets, model weights and software releases dating back to 2025.
The company also compares Laya with TypeSafe AI’s Jev, claiming faster execution under its stated testing conditions. Those performance comparisons, along with claims about language coverage, calibration and benchmark results, have not been independently verified. The supplied material does not include complete benchmark tables, methodology or hardware specifications.
