Loggi’s Rio deliveries, at real scale
Loggi is one of Brazil’s largest last-mile carriers. Its open benchmark (loggibud, MIT-licensed) models real urban delivery from public geography. The source went offline — repository and dataset both gone — so this Rio de Janeiro instance was recovered from the Internet Archive and committed with its license. 1,765 deliveries, one depot: a genuinely large problem, solved.
1,765 deliveries across one city
Each blue point below is a real delivery location in Rio de Janeiro; the red marker is the distribution hub. With a vehicle capacity of 180 and total demand of 9,773, this is the kind of instance a real operator faces every morning — far beyond anything an exact solver can touch.
Exact can’t prove twelve; the engine solves 1,765
Take a twelve-stop slice. HiGHS and BIS’s HGS agree on 31.92 km — but here the exact solver cannot even prove that is optimal within the time budget: Loggi’s loose capacity makes the MILP relaxation weak, so the proof is hard on a problem the metaheuristic finds trivial. Scale that contrast to the full instance and the exact approach is simply not in the conversation.
On all 1,765 deliveries, PyVRP — the reference open-source Hybrid Genetic Search — returns a 1,140.7 km solution with 55 vehicles in about 48 seconds. There is no published best-known for this instance, so we report the metaheuristic cost and the exact-on-a-slice agreement rather than claim optimality.
A fleet problem this size — or bigger?
BIS builds routing engines that run in your stack, at your scale, on your real geography. Let’s scope it.