Freight Intelligence, Built by Neblo

Neblo processes 300,000+ pieces of freight information each month. Today, most of that work runs through off-the-shelf frontier AI models.

We are excited to announce the rollout of our own freight-trained ML models. On the work the model selected to handle in held-out testing, it matched the frontier-AI teacher on every item. The same work ran with far lower latency and reduced how much traffic needed the frontier model.

Much lower latency

With frontier AI normalized to 100%, Neblo's freight model uses about 1% of the processing time on selected work. That is a 99% reduction in processing latency for the work the model handles.

Relative processing time

Frontier AI = 100%; Neblo freight model = 1%; lower is better.

Cost per task

With frontier AI normalized to 100%, Neblo's freight model costs 16% as much per task for the same routine work using standard AWS on-demand hosting. That is an 84% cost reduction per task.

Cost per task

Frontier AI = 100%; Neblo freight model = 16%; lower is better.

Trained with a teacher-student approach

Neblo has a large and growing body of high-quality freight data. We use frontier AI as the teacher, then train smaller freight-focused student models on that data to recognize routine work.

This gives Neblo broad frontier reasoning and specialized freight intelligence trained on real operating data.