The English root uses ModernBERT-large (421M parameters, 512-token context).
Apache 2.0 · 100+ languages
Laya AI: calibrated decisions in one pass
Built for text, email, tickets, and JSON. Ask typed questions and get calibrated probabilities—without generated prose.
TextEmailTicketsJSON
Decision outputs, not essays
Keep the input state separate from the answer space. Define questions and options at request time; Laya AI scores them in one pass and returns calibrated probabilities you can gate, log, reproduce, and audit.
The multilingual model uses mmBERT-base (322M parameters, 100+ languages).
RLCD and strictly proper scoring rules reward calibrated probability reporting.
The English, multilingual, and typed-decisions weights are Apache 2.0.
Published evaluation: 0.732 email triage, 0.883 fact checking, and 0.991 intent routing.
Built for evidence-backed decisions
Use Laya AI where a defensible probability distribution matters more than fluent generation.
Typed answers
Ask choice, scoring, or yes/no questions. The answer space is defined at request time, without retraining the encoder.
One-pass batching
Parallel checks—such as urgency, department, refund intent, and churn risk—run in the same forward pass.
Automatic routing
Router detects language and script, chooses the checkpoint, and supports an explicit language override.
Calibrated probabilities
Probability distributions support gating, sampling, and escalation. Fit temperature on your own data first.
No generated prose
No free-form replies means less parsing and less hallucination. Each output stays tied to the supplied options.
From input to decision
The documented path is short: install the package, define the state and answer space, call the Router, then act on returned probabilities.
Install Laya AI
Install the laya package from PyPI and load only the checkpoint your service needs.
Define the state
Provide an email, ticket, document, or JSON payload. Put decision rules in question instructions and option descriptions.
Ask typed questions
Combine department, urgency, and refund-intent questions in the same call.
Route and act
Read the selected option and probability, log routing metadata, then apply your threshold.
Evaluate before rollout
Label live samples, compare with your current policy, fit temperature, then set escalation thresholds.
Measured results and known limits
These figures come from Laya’s project documentation. Each result identifies the checkpoint, workload, and hardware used for the measurement.
Fine-tuned checkpoint across 2,000 decisions.
Across 24,314 evaluated questions.
After calibration on the published evaluation.
Recorded by the project’s local evaluation.
Measured conditions
Figures identify the checkpoint, workload, and hardware used for the measurement.
Known trade-offs
Base checkpoints score near chance on zero-shot typed decisions, ordinal scoring is the weakest primitive, and a default 77-option Banking77 run falls to 0.425 because option tokens are tightly budgeted.
Laya AI vs. TypeSafe Jev
A compact comparison against published TypeSafe Jev figures.
| Metric | TypeSafe Jev 1.13.0 | Laya AI, routed |
|---|---|---|
| Typed decisions, 2,000 | 0.727 | 0.766 |
| AG News, 4 labels | 0.910 | 0.950 |
| DAIR Emotion, 6 labels | 0.480 | 0.595 |
| Banking77, 72 vs. 77 labels | 0.870 | 0.425 |
| ECE, lower is better | 0.246 | 0.081 |
| p50 latency, 1 question | 236–276 ms | 32.8 ms |
| Languages usable | No published benchmark | 45 of 51 |
| Weights | Closed API | Apache 2.0 |
TypeSafe Jev figures come from its published documentation; benchmark sample sizes and prompts differ from Laya AI’s routed runs.
High-option and soft-distribution results
- Banking77: TypeSafe Jev 0.870; Laya AI 0.425 on 77-option questions.
- Typed-decisions soft accuracy: TypeSafe Jev 0.580; Laya AI 0.471.
- Raw ECE: TypeSafe Jev 0.144; Laya AI 0.213 before temperature fitting and 0.081 after domain temperature fitting.
Self-host the open weights
Laya AI ships under Apache 2.0 and includes a documented server path. The weights, requests, logs, and threshold policy can stay inside your own environment.
Install and serve
Install laya with its serve extra, then run laya-serve. Environment variables control device and preload behavior.
Jev-compatible endpoint
POST /v1/systemone follows the TypeSafe Jev shape, so existing clients can change only their base URL.
Control access
The server binds to 0.0.0.0 without authentication. Set LAYA_API_KEY before exposing it beyond a private network.
Operate with limits in mind
Keep English on its root model, route other input to the multilingual model, monitor option length, and reserve Laya AI for calibrated decisions.
Language routing and coverage
Laya AI documents model capability and routing evidence separately. The multilingual checkpoint supports 100+ languages; the shared benchmark measured 45 of 51 languages.
- English uses the ModernBERT-large root checkpoint.
- Non-Latin scripts route to Laya AI’s multilingual checkpoint.
- Script detection takes under 0.5 milliseconds.
- Shared benchmarks found 45 of 51 languages usable.
- Version 0.3.7 improved plain-ASCII Romance languages and Brazilian Portuguese.
- Override detection with lang_guess or preload up to two checkpoints.
Laya AI FAQ
What is Laya AI?
Laya AI is a family of non-autoregressive decision models. It reads an input state and typed questions, then returns structured answers with probability distributions instead of free-form text.
Which checkpoints are available?
The hub includes a 421M-parameter English root, a 322M-parameter multilingual checkpoint, and a 421M-parameter typed-decisions checkpoint. Each is documented with its encoder, context length, and appropriate workload.
Can Laya AI replace a language model that writes replies?
No. Laya AI is a System 1 decision model. Use it for choices, scores, routing, guardrails, moderation, and confidence signals; hand generated replies to another component when the product genuinely needs them.
How should probabilities be used safely?
Fit temperature on your own data, inspect calibration, and set conservative thresholds. The published evaluation reports 0.753 overall in-task accuracy and 0.947 accuracy at 50% coverage after calibration.
What are the zero-shot limits?
Base checkpoints score near chance zero-shot. High-option questions also strain the token budget. Evaluate your own workload before deployment.
Can I self-host Laya AI commercially?
The checkpoints are Apache 2.0. Review the license with your compliance process and enable authentication outside a private network.
What did the published evaluation measure?
It covers in-task and held-out zero-shot families, calibration, and latency. Overall zero-shot accuracy is 0.651; the larger in-task set reports 0.753.