Convai ships Laya, a 421M ModernBERT decision model, Apache 2.0

Convai Innovations has posted Laya, a decision head built on ModernBERT-large that returns single-forward-pass classifications in about 39.5ms for the 421M English checkpoint and 32.8ms for the 322M multilingual variant on a Tesla T4. It ships under Apache 2.0.

The company posts three checkpoints: an English base (ModernBERT-large, 421M, 512 context), a multilingual variant on mmBERT-base (322M, 1024 context, 100+ languages), and a typed-decisions checkpoint fine-tuned on that benchmark's own training split. On its head-to-head numbers, Laya's typed-decisions checkpoint hits 0.766 accuracy against Jev 1.13.0's 0.727, with p50 latency of 32.8ms versus Jev's 236 to 276ms, a 7.8x speedup the card attributes to non-autoregressive scoring.

The model card is blunt about what that headline number does not represent. On zero-shot typed decisions the card reports Laya at 0.362 and the multilingual variant at 0.352, against a 0.318 random baseline and a 0.461 majority-class baseline. "The 0.766 belongs to the checkpoint fine-tuned on that benchmark's own training split. Laya is a fast base to specialise, not a zero-shot decision engine," the card reads. Calibration ships poor as well: refitting one temperature per question type moves mean ECE from 0.466 to 0.081.

Convai frames the target workloads as invoice processing, ticket routing, moderation, guardrails and agent-trace observability, the classification layer around an LLM stack rather than the LLM itself. The multilingual variant's macro-average on MASSIVE Intent across 51 languages is 0.227, so the 100-plus language coverage buys reach, not accuracy.