Terra Quantum is opening Early Access to TQrouting, our vehicle routing optimization engine that set 27 new best-known solutions in the CVRPLib BKS Challenge, on a fraction of the compute reported by the winning entry. Places are limited and include access to support from the team that built it.
In last-mile delivery, the planning window is short by nature. Orders lock late. Vehicles load early. There are typically only twenty or thirty minutes in between, and whatever plan exists when that window closes is the plan the fleet runs all day.
Route quality decides driver hours and vehicle count, which means it decides margin. So the planning problem is not really "find good routes." It is "find good routes, for this many stops, under all of these constraints, before the doors shut."
That is where most optimization software runs out of room. As stop counts rise and constraints multiply, general-purpose solvers do not degrade gently. On the largest problems they stop finding a feasible plan at all.
That is what TQrouting was built for.
What TQrouting is
TQrouting is an embeddable vehicle routing optimization engine. It combines in-house-developed, ML-based decomposition with advanced mathematical optimization and intelligent parallelization to produce near-optimal, dispatch-ready plans for problems with tens of thousands of stops, inside tight operational windows.
At scale, decomposition is the hard part: breaking a very large problem into pieces that can each be solved well, without losing quality where the pieces meet. That is the part we went after.
It runs as a library inside your own environment, on-premises or in your private cloud, so your data does not leave it. It sits underneath the logistics, dispatch and transport management systems you already run, as the optimization component, where scale and constraint density have become limiting.
"Most deep tech is a promise with a date ten years out. TQrouting set 27 records in a public challenge anyone can verify, and it can run on infrastructure our customers already own. We moved what is possible at the top end, and the top end is where the world's freight actually moves." - Markus Pflitsch, CEO and Chairman, Terra Quantum
What the benchmarks show
The CVRPLIB BKS Challenge
Earlier this year, CVRPLib, a widely recognized public benchmark library for vehicle routing optimization, ran a 30-day open challenge on a new set of 100 large-scale instances of 1,000 to 10,000 customers. It was organized by a group of distinguished vehicle routing researchers: Eduardo Queiroga, Rafael Martinelli, Anand Subramanian, Eduardo Uchoa and Thibaut Vidal, and was open to teams from academia and industry. Nineteen teams were approved to compete. The instances were new, so every team started under the same conditions.
The field included industry-academia collaborations between Huawei Technologies and City University of Hong Kong, ByteDance and Southern University of Science and Technology, and Google with the universities of Calabria and Bologna.
TQrouting set 27 of the final best-known solutions. Our results concentrated at the top of the size range: 7 of the 10 largest instances, including all three in the 9,570 to 10,000 stop range. Results were verified by the academic community, and every team's methods and computational resources are published on the challenge site.
The organizers' post-competition analysis of computational effort is worth reading alongside the results. Terra Quantum used approximately 5,200 core-days. The first-placed entry reported approximately 42,600 CPU-days on Huawei's internal cluster, more than eight times more.
"Growing from one thousand to ten thousand customers does not make a route optimization problem ten times harder. The number of possible solutions explodes combinatorially. At that scale, standard optimization solvers do not simply get slower, but often fail to find even a feasible solution within an operational planning window. This requires a new optimization engine designed to specifically handle this complexity. We built TQrouting for that purpose." - Dr Giorgi Tadumadze, Lead Scientist, Operations Research, Terra Quantum
Our own published benchmark studies
Separately, we have published gap-to-best-known results on various popular benchmark datasets, measured under fixed runtime limits against commercial, open-source and state-of-the-art algorithms. Public instances, reproducible by anyone, run by us:
- 0.40% average gap to best-known on CVRP up to 1,000 customers, within one minute
- 1.68% on CVRPTW, 0.022% on MDVRP, 0.02% on TOP, each within one minute
- 3.10% at 10 minutes and 2.42% at 60 minutes on instances of up to 30,000 customers
The full methodology, instance sets and comparison figures are in the benchmark articles linked at the end.
Early Access is open
Early Access places are limited and allocated on fit, as each one comes with direct support from the Operations Research team. If your problem is the kind TQrouting was built for, a place gets you:
One month, no cost. A time-boxed license to run TQrouting on your own data, inside your own infrastructure, on the problems you actually plan.
Direct support from the people who built it. Installation support, documentation, and regular check-ins with our team.
Who it is for
- Optimization, operations research and data science teams at large-scale logistics operators in parcel, postal, e-grocery and last-mile delivery
- Software providers and ISVs building routing into their own products, who need a high-performance optimization core underneath
- Teams running VRP on general-purpose or open-source solvers and hitting limits as instances grow
Apply
Applications are open at https://terraquantum.swiss/vehicle-routing-optimization-engine/
If your problem is a good fit, we will follow up with a short call to cover problem size, the constraints that matter, what you run today, and what better would need to look like. From there we scope the evaluation and set up your month.
If you would rather see it running first, request a demo.
Read the benchmark work: