Experiential Labs launches open-source WMO Serve to route agent tasks to distilled small models at lower cost

Experiential Labs introduced world-model-optimizer’s wmo serve, an open-source tool that uses existing agent traces to improve specialized models through distillation, routing, and token compaction. It runs a local OpenAI-compatible endpoint that sends simpler tasks to smaller custom models and harder ones to frontier models, with a hosted option promising 40%+ lower costs.

The launch reflects a growing push to reduce AI inference costs by combining model routing with continual fine-tuning from production traces. If these systems can preserve quality while automating optimization, more agent workloads may shift from always-on frontier models to cheaper hybrid architectures.