Laya

Developers can use it to run local choice, ordinal-score, and boolean decisions, or route requests among Core ML model bundles.

Availability

Install from PyPI; model bundles available on Hugging Face

Pricing model

Apache-2.0

Deployment

Local inference on Apple Silicon; download once, then run offline

Privacy

Subsequent predictions stay local

Licence

Apache-2.0

• Laya returns typed answers with calibrated probabilities in a single forward pass. ^[1]

  **Multilingual, non-autoregressive System 1 decision model.** Give it a **state** (text, email, ticket, or JSON) and **typed questions**; it returns typed answers with mathematically calibrated probabilities in a single forward pass (~33 ms) across 100+ languages.

• The English checkpoint uses ModernBERT-large and has 421M parameters. ^[1]

  The English checkpoint is at the repo root; the other two are bundled subfolders, and only the one you request is downloaded:

• The package can be installed with pip install laya. ^[1]
• Version 0.4.0 is available as the laya-coreml Python package on PyPI.
• The runtime supports batch and long-text prediction, hooks, structured decisions, calibration, HTTP/MCP, evaluation, and Router behavior. ^[2]

  prediction, hooks, structured decisions, calibration, HTTP/MCP, evaluation and Router behavior now

• After downloading a model, users can run inference offline on Apple Silicon. ^[2]
• The runtime returns probabilities for choice, ordinal-score, and boolean questions without autoregressive decoding. ^[2]

  Laya returns probabilities for **choice**, ordinal **score**, and boolean **noul**

• Subsequent predictions stay local. ^[2]

  models download before initialization; subsequent predictions stay local.

• The project is an independent port and is not an official Convai Innovations or Apple release. ^[2]

  Not an official Convai Innovations or Apple release.