421M browser-based encoder outperforms 1.5B LLM in agent decision-making tasks

Developer Vishal Mysore built layaAgent, a browser-based AI agent that uses a 421-million-parameter typed-decision model called Laya to handle routine decisions without a server or API key. Laya, built on ModernBERT-large and exported to ONNX for in-browser use, answers only structured yes/no or multiple-choice questions, making its decisions fully inspectable as probability scores. A larger 1.5B-parameter LLM, Qwen2.5-1.5B, is invoked only when Laya's confidence falls below a set threshold, reducing reliance on slower and costlier generative models. In testing, Laya correctly selected tool arguments 71% of the time overall, rising to 90% accuracy for fixed-list arguments, with the extractor supplying the correct span 91% of the time. The project is open-source under the Apache 2.0 license, with a live demo and adjustable confidence thresholds that instantly reroute recorded decisions without rerunning any model.