The Pulse
Convai Innovations' Laya Model Returns Typed Decisions in One Pass
Convai Innovations' Laya model uses ModernBERT and a decision head to return typed answers, probabilities and confidence values for routing, scoring and calibrated classification workflows.
AI.info Team ·
Laya Treats AI Output as a Decision, Not a Paragraph
Convai Innovations' Laya is a 421-million-parameter model that accepts text, email, tickets or JSON as state and answers typed questions without generating a natural-language response. The model card describes it as a multilingual, non-autoregressive decision model that returns typed answers with probabilities in one forward pass.
Laya is available through the Convai Innovations model repository on Hugging Face under the Apache 2.0 license. The repository holds three checkpoints: the English model at the root, a multilingual model in a subfolder, and a typed-decisions model in a second subfolder. The English and typed-decisions checkpoints have 421 million parameters, while the multilingual checkpoint has 322 million.
The model supports three question types. choice returns a selected option, a probability for each option and confidence. score returns an expected level on an ordinal rubric, along with a distribution and confidence. noul returns a calibrated probability that a statement is true.
ModernBERT Supplies the Encoder
The English and typed-decision checkpoints use ModernBERT-large as their backbone. The architecture combines the 395-million-parameter encoder with a decision head trained from scratch, producing 421 million total parameters.
Laya scores each answer option at its own [MASK] marker before applying a softmax across the options for that question. The standard English checkpoint has a 512-token budget per question. The typed-decision checkpoint is listed with a 1,024-token context window.
Because the model returns structured outputs directly, an application does not need to ask a generative model to emit JSON and then parse a paragraph. Laya returns the selected answer, its probability distribution and a confidence value. The repository says that it never generates text.
One Forward Pass for Routing and Triage
The quickstart demonstrates support-ticket routing, urgency scoring and churn-risk detection in one call. A sample ticket can be assigned to a department, rated for urgency and checked for whether the user threatens to cancel. The repository also provides a router that selects among the English, multilingual and typed-decisions checkpoints.
On a Tesla T4, the model card reports 39.5 milliseconds for one question with the English checkpoint and 32.8 milliseconds with the multilingual checkpoint. Ten questions take 158.6 milliseconds and 72.3 milliseconds, respectively. For 50 questions, the reported times are 771 milliseconds for English and 337 milliseconds for multilingual. The documentation reports 103 to 332 questions per second when requests are batched.
The project exposes a Python package through PyPI. Developers can install it with pip install laya, load a checkpoint, provide a state object and submit several typed questions in one call. The repository also links to a browser-based Laya demo.
Benchmark Results Vary by Checkpoint and Task
In a typed-decisions benchmark covering 400 cases and 2,000 decisions, the fine-tuned laya-typed-decisions checkpoint records 0.766 accuracy, 0.471 soft accuracy, a 0.062 Brier score, 0.213 expected calibration error and a 0.242 score mean absolute error. The same table lists 0.362 accuracy for the English checkpoint and 0.342 for the multilingual checkpoint.
The fine-tuned checkpoint records accuracy of 0.804 on invoice processing, 0.766 on security incidents, 0.764 on customer service and 0.730 on agent-trace observability. By question primitive, the results are 0.857 for noul, 0.733 for choice and 0.723 for score.
On English tasks, the model card lists 0.947 accuracy on AG News, 0.830 on BoolQ, 0.573 on DAIR Emotion, 0.698 on prompt-injection examples and 0.372 on the ordinal SST-5 task. The multilingual checkpoint reaches a macro average of 0.227 with a macro expected calibration error of 0.733 across 51 languages on the MASSIVE intent benchmark.
Limits and Intended Use
The documentation warns that the base English and multilingual checkpoints perform near chance on the typed-decisions benchmark without task-specific fine-tuning. It identifies 0.766 as the result for the checkpoint fine-tuned on that benchmark's training split, rather than a zero-shot result.
The model card recommends keeping choice questions below about 20 options because the fixed label budget can reduce accuracy when too many options are supplied. It also identifies ordinal score questions as the weakest primitive in the listed English results. The confidence score does not warn when an input is unreadable, so the repository says the choice of checkpoint must be made before the forward pass, which is the purpose of the router.
Laya is presented as a decision component for applications that need bounded outputs, probability estimates and local deployment. Its results depend on the selected checkpoint, the task format and whether the model has been fine-tuned for the target workflow.