Laya

Laya is built by GitHub user NandhaKishorM, and the README also credits Convai Innovations. It is a multilingual "System 1" decision model with an SDK: you give it some text and a few typed questions, and in a single non-autoregressive forward pass it returns a choice, a score or a null answer, each with probabilities. It doesn't generate a reply token by token the way a large language model does, so it suits agents that need to pick from a fixed set of options quickly. You can call it from Python, the command line, HTTP or MCP, and there are TypeScript, Node, browser and C++ ONNX versions. The author says the model covers more than 100 languages, and the speed and accuracy figures in the README come from the author's own benchmarks.

Who it's for:developers building AI agents who need a fast classification or decision step. Not useful if you don't write code.

Before you start: you need Python 3.10 or newer and install it with pip install laya. The first run downloads a model checkpoint; once it is saved locally, Laya can run offline. Licensed under Apache-2.0. Based on the project's public documentation. Not tested by AI Inspiration World.