v1.14.1

Added

• Laya decision models (convaiinnovations/laya, Apache-2.0).
  Laya is a System One model: give it a state and typed questions (noul, choice, score), and it answers with
  calibrated probabilities in one forward pass, in 100+ languages.
  • A new Laya engine runs it (Engines page; Windows, macOS and Linux, on CPU, NVIDIA or Apple GPU).
  • POST /v1/systemone and the playground answer with a Laya model exactly as they do with a Jev one. The response
    adds routing, which names the checkpoint that answered (English or multilingual).
  • A Laya model always loads on the Laya engine, beside your chat model, whichever engine is active. Chat stays
    available, and the Workspace sidebar links to the playground while it is loaded.
  • Search for "laya" in Discover to find it. Its download fetches the English and multilingual checkpoints
    (~1.5 GB).
  • The GGUF conversions of Laya on Hugging Face use a separate runtime (ggmlc) that llama.cpp can't load. They are
    marked as not loadable and point you at the model that runs.

Fixed

• A model loaded in its own slot beside the chat model (an embedding model, for example) can no longer be deleted
  while it runs. Delete now asks you to eject it first, as it already did for the chat model. Before, the check
  only looked at the chat model's slot.