nandhakishorm/laya is a non-autoregressive decision engine that, based on text, picks among options, computes scores, and returns yes/no judgments in a single forward pass. The latest release is v0.4.9. The provided materials do not say what changed in this version. The repository is written in Python and uses the Apache-2.0 license.
Instead of continuing to generate sentences, its core is producing judgment results of a predefined type. According to the project description, it supports more than 100 languages and includes a router that selects the checkpoint matching the request. The README summary explains that it was trained with reinforcement learning aligned to a strictly appropriate scoring rule (RLCD). It also gives a figure of p50 32.8 milliseconds per question on a T4, but that is its own measurement. So rather than replacing generative models, it is closer to a tool for tasks with a fixed answer form, such as selection or classification.
The limitations stated in the README are clear as well. When input exceeds 4,000 tokens, accuracy fluctuates between 8 and 17 out of 20 questions, and on the MASSIVE benchmark the macro accuracy for multilingual models is 0.40. The 0.766 for typed-decisions is the figure for a fine-tuned checkpoint with no committed results file. It is therefore difficult to judge long inputs or multilingual performance from these numbers alone.