Laya turns typed LLM decisions into a single pass

Convai Innovations has released Laya, an Apache 2.0 open-weights model for bounded decisions such as routing tickets, scoring urgency, and flagging content. The project reports 32.8 ms median latency for one question, roughly 6 to 8 times faster than the published 236 to 276 ms range for TypeSafe Jev, a closed-API decision model.

Laya’s “System 1” label describes a fast, non-autoregressive encoder that accepts a state, such as text, email, ticket data, or JSON, alongside typed questions. It scores the supplied answer options and returns typed answers with probability distributions in one forward pass across more than 100 languages. Direct option scoring avoids token generation, malformed generated JSON, and free-form answers outside the schema, although the model can still misclassify an input.

Three checkpoints, one router

The release packages three fine-tuned encoders, with only the requested checkpoint downloaded:

• convaiinnovations/laya: ModernBERT-large with 421 million parameters, a 512-token context window, and training focused on English text, guardrails, and email triage.
• laya-multilingual: mmBERT-base with 322 million parameters, a 1,024-token context window expandable to 8,000 tokens, and coverage for more than 100 languages. The release reports about 2.2 times the speed of the larger checkpoint.
• laya-typed-decisions: ModernBERT-large with 421 million parameters, fine-tuned for four typed-decision workflows. It records 0.766 accuracy on the project’s benchmark.

The built-in Router detects language and script in under a millisecond, then selects the appropriate checkpoint before inference. That routing step prevents severe failures on unsupported scripts: the English checkpoint recorded 0.000 accuracy with 0.952 confidence on Khmer in the published evaluation. A confidence threshold would accept many of those incorrect predictions, while script-aware routing sends them to the multilingual model.

RLCD rewards honest probabilities