Laya
GPU · public betaA hosted instance of Laya, an open-source non-autoregressive decision model. Send it a state (text, email, ticket, JSON) plus typed questions and get back typed answers with calibrated probabilities in a single forward pass — no text generation, nothing to parse.
Laya is an open-source alternative to Jev, the proprietary "System One" model: the same idea (choice, score and noul questions with calibrated confidence, evaluated in parallel against one state), but Apache-2.0 weights you can self-host, no per-token cost and no data leaving your infrastructure.
Laya vs Jev
| Jev | Laya (this endpoint) | |
|---|---|---|
| Latency (1 question) | ~100 ms typical (70–500 ms) | ~35–40 ms on GPU; 10 questions batched in ~160 ms |
| Question types | choice, score, noul | Same three, same request/response shape |
| Architecture | Undisclosed | ModernBERT-large / mmBERT encoder + decision head, 322–421M params |
| Calibration | RLCD, details not public | RLCD with published proper-scoring-rule reward; calibration curves in repo |
| Languages | Not published | English model + 100-language model, auto-routed per request |
| Limits | Up to 255 options per choice | 512 tokens/question (English), 1,024 (multilingual) |
Full benchmarks and calibration plots: Laya README · model card.
Endpoint
POST https://laya.pensero.ai/predict (alias: https://jev.pensero.ai/predict) Content-Type: application/json
| Field | Type | Description |
|---|---|---|
state |
object | string | What to decide about. |
questions |
object | Map of question id → {type, instructions, criteria}. Types: choice, score, noul (boolean). |
model |
string, optional | english, multilingual or typed-decisions. Auto-routed by language when omitted. |
Example
curl -s https://laya.pensero.ai/predict \
-H 'content-type: application/json' \
-d '{
"state": {"subject": "Duplicate charge on invoice #4411",
"body": "We were billed twice for March. Refund today or we cancel."},
"questions": {
"department": {"type": "choice", "instructions": "Which department should handle this?",
"criteria": {"billing": "invoices, refunds", "technical": "bugs, outages", "sales": "pricing"}},
"urgency": {"type": "score", "instructions": "How urgent is this?",
"criteria": ["not urgent", "soon", "critical"]},
"churn_risk": {"type": "noul", "instructions": "Does the user threaten to cancel?"}
}
}'
Response: answers.department.choice, per-option probabilities, a calibrated confidence, and routing metadata explaining which checkpoint was used.
Limits
- Rate limit:
60/minuteper client IP. Exceeding it returns429; check theX-RateLimit-*headers. - Context: 512 tokens per question (English), 1,024 (multilingual / typed-decisions).
- No authentication, no persistence — requests are not stored.
Interactive OpenAPI docs: /docs