After reading this, the reader knows how Laya differs from a generative model and which performance claims remain vendor-tested.

Laya Releases an Open Decision Model

The 421-million-parameter model returns bounded choices and probabilities instead of prose, targeting routing and scoring tasks that do not need free-form generation.

Convai Innovations has released Laya, an Apache-2.0 model designed to make typed decisions from text or structured input. The model accepts a state, a question and allowed answers, then returns a choice, ordinal score or boolean probability with confidence information.

In plain terms, Laya does not write an explanation or invent an answer outside the choices supplied by the application. It evaluates the options in one forward pass. That makes it a specialized component for tasks such as ticket routing, moderation gates and email triage rather than a replacement for a general chatbot.

Why it matters

Many production workflows use a generative model for a task that is fundamentally classification. The application asks for JSON, waits while the model generates tokens, parses the result and handles malformed output. A model built to score permitted choices directly can remove those steps and make the output space explicit.

That does not eliminate model error. Laya can still select the wrong option or report a poorly calibrated probability. Its narrower interface instead changes the failure mode: the system cannot produce an unexpected prose answer when the application asked it to choose A, B or C.

The model card describes a 395-million-parameter ModernBERT-large backbone plus a decision head, bringing the total to 421 million parameters. Each question, its options and the supplied state share a 512-token input budget. Multiple questions can be evaluated in one batch.

Convai Innovations says it trained the model through Reinforcement Learning for Calibrated Decisions. The reward uses scoring rules intended to favor probability distributions that match observed outcomes. Multi-turn examples use temporal-difference learning across prefixes so later outcomes are not exposed to earlier decisions.

Benchmarks need outside reproduction

The published comparison reports 83.8% in-task macro accuracy for Laya against 67.8% across four workflows for Jev, a commercial decision model. It also reports median latency of 38.4 milliseconds for one question. These figures were produced by Laya’s developer, and the compared systems do not share a fully documented public evaluation setup.

The model card itself signals several limitations. Its benchmark is mainly English, most inputs are under 512 tokens, and performance is sensitive to the available choices. A developer can obtain a confident answer while omitting the correct option. Distribution shift can also break calibration even when historical evaluation looked reliable.

The claimed zero self-hosted inference cost should be read as zero usage fee, not zero operating cost. Local deployment still consumes hardware, electricity and engineering time. The practical comparison is total cost per reviewed decision at an acceptable error rate.

The release is nevertheless testable. Weights and code are public, so independent users can run the same checkpoint on their own cases, record accuracy and calibration, and compare direct decision scoring with a generative baseline. Useful tests should include an abstain option and measure outcomes after human escalation.

The next evidence should come from evaluations outside the training domains, with fixed datasets, hardware and baselines. Laya’s most important contribution is not a benchmark lead. It is a concrete alternative to using free-form generation for every AI-assisted decision.

Verification

  • Tier 0 — VERIFIED: The Laya checkpoint, model card and weights were published on Hugging Face under Apache 2.0, with repository activity inside the reporting window. Primary source: https://huggingface.co/convaiinnovations/laya
  • Tier 0 — VERIFIED: The model card specifies 421 million parameters, typed outputs and a 512-token input budget per question. Same primary source.
  • Tier 1 — DEVELOPER-REPORTED: Accuracy, latency, calibration and Jev comparison figures were published by Convai Innovations and were not independently reproduced here. Same primary source.
  • Tier 2 — ANALYSIS: Deployment, failure-mode and evaluation implications are editorial analysis.

Glossary candidates

  • Calibration: How closely predicted probabilities match observed frequencies.
  • Non-autoregressive model: A model that produces its result without generating an output sequence token by token.

Cold-reader sentence: Laya turns bounded options into probabilities without generating prose, but its reported accuracy and speed still need independent testing.