laya

Laya is a family of open decision models by Convai Innovations, released under the Apache-2.0 license. A decision model reads a state — a message, an email, a support ticket, a JSON object — together with a set of typed questions, and returns a typed answer with calibrated probabilities for every question in a single forward pass. It never generates text.

Models

Each model carries an fp16 and an fp32 graph over one weights file. It loads fp16 on a CUDA GPU and fp32 on the CPU; the -fp16 and -fp32 tags pin one, for example laya:en-fp32. laya routes Turkish and other non-English text to laya:multilingual, and pulling it pulls both targets.

Usage

ollaya run laya --preset triage "I was charged twice for my subscription this month and want a refund."

Or call the local API:

curl http://localhost:11435/api/decide -d '{
  "model": "laya",
  "state": "I was charged twice for my subscription this month. Please refund the second charge.",
  "questions": {
    "department": {
      "type": "choice",
      "instructions": "Which team should handle this ticket?",
      "criteria": {
        "billing": "Payments, invoices and refunds",
        "technical": "Bugs, errors and outages",
        "account": "Login, profile and settings"
      }
    },
    "urgency": {
      "type": "score",
      "instructions": "How urgent is this ticket?",
      "criteria": ["Can wait", "Needs attention this week", "Needs attention today"]
    },
    "refund": {
      "type": "noul",
      "instructions": "The customer asks for money back."
    }
  }
}'

{
  "model": "laya:en",
  "answers": {
    "department": {
      "type": "choice",
      "choice": "billing",
      "confidence": 0.7781,
      "probabilities": {"billing": 0.8521, "technical": 0.0611, "account": 0.0868}
    },
    "urgency": {
      "type": "score",
      "score": 1.1982,
      "confidence": 0.3418,
      "legend": {"0": "Can wait", "1": "Needs attention this week", "2": "Needs attention today"},
      "probabilities": {"0": 0.1203, "1": 0.5612, "2": 0.3185}
    },
    "refund": {"type": "noul", "noul": 0.9127}
  },
  "usage": {"input_tokens": 118, "output_tokens": 0},
  "routing": {"router": "laya:latest", "model": "laya:en", "route": "english", "reason": "English Latin text"},
  "state_truncated": false,
  "done_reason": "decide",
  "created_at": "2026-09-24T09:30:12.418Z",
  "total_duration": 18734512,
  "load_duration": 0,
  "eval_duration": 16302117
}

model is the checkpoint that answered, and routing says why. See the API reference for every field, and TypeSafe compatibility for /v1/systemone.

Question types

confidence is the normalized top probability, (K · p_max − 1) / (K − 1) for K options: 0 when every option is equally likely, 1 when one option has all the probability. Every option needs room in the context: about 125 options fit laya:en, and 250 fit laya:multilingual.

Performance

On an RTX 4090 at fp16, a request with five questions takes 16.3 ms on laya:en and 9.1 ms on laya:multilingual (median, Ollaya's Phase 0 benchmark).

Figures published on the Laya model card, measured on an NVIDIA Tesla T4:

On batched calls laya:multilingual is up to ~2.2× faster than laya:en; for a single question the two are close. laya:typed-decisions reaches 0.766 accuracy on typed-decisions, against 0.727 published for Jev 1.13. Jev latencies come from AbdelStark/jev-benchmarks and nibzard/decision-model-benchmark; setups differ.

ONNX export

Ollaya runs Laya as ONNX. Across 2,383 questions per checkpoint (en, multilingual and typed-decisions), the ONNX export chose the same answer as the PyTorch fp32 reference 100% of the time, with a maximum probability difference of 1.1 × 10⁻⁴. The fp16 graph, used on CUDA GPUs by default, can differ from fp32 on near-ties.

Limitations

• Zero-shot typed decisions. The base checkpoints (en, multilingual) are near chance zero-shot on typed-decisions (0.362). Use laya:typed-decisions for those workflows.
• Many options. Choice questions with more than ~20 options are weaker: 0.425 on Banking77, against 0.870 for Jev.
• Calibration. The raw checkpoints are over-confident unless their temperatures are refit, and laya:multilingual ships without refit temperatures. For thresholds you rely on, refit calibration on your own labelled data and bake it in with CALIBRATION in a Modelfile.

Weights and license

Apache-2.0. Laya is developed by Convai Innovations — see the model card on Hugging Face. Ollaya downloads the weights from that repository, pinned to a commit and checked against sha256; it never re-hosts them.