Yandex Routing and RouteQ: What the Documentation Says

Yandex Routing is now documented as RouteQ. How its vehicle routing model, penalty cost, and probabilistic search work, with an illustrative cost example.

Last verified: October 20266 min read

Yandex Routing is a cloud vehicle routing service from the Russian technology company Yandex, and since at least 2022 it has been marketed under the name RouteQ. Its public documentation describes a planner that takes one depot, a fleet, and a list of orders with time windows, then searches for the plan with the lowest combined cost of resources and rule violations. The search is probabilistic and time-capped, so it returns a good plan, not a proven optimum, and the same input can produce different routes.

Key takeaways

  • A February 21, 2022 press release calls RouteQ the former Yandex.Routing and a Yandex subsidiary. No newer ownership news was found for this article.
  • Each request has one depot. Vehicles carry capacities, and orders carry coordinates, time windows, and service durations.
  • Hard rules, such as vehicle capacity and order incompatibility, cannot be broken. Soft rules, such as time windows, add penalties to a single cost number.
  • The documented search starts from a first route, keeps improvements, and sometimes accepts a worse route with a probability that falls as the route worsens.
  • Penalty weights decide which plan wins, so two services can disagree on the best plan without either being wrong.

The product and its name

A press release distributed on February 21, 2022 with a Moscow dateline describes RouteQ as formerly known as Yandex.Routing and as a subsidiary of Yandex that provides cloud-based delivery management software. The Yandex Routing address on the Yandex site now redirects to a RouteQ documentation site, which fits that description. The same release quotes efficiency gains from earlier case studies, such as 20 to 30 percent better delivery efficiency. Those figures come from the company and no independent measurement was found. Whether the service can be bought or used from Canada or the United States is a compliance question that this article does not assess.

What the documentation says it models

The service documentation frames the task as a vehicle routing problem, the multi-vehicle relative of the traveling salesman problem. A request holds a single depot with working hours, a list of vehicles or couriers, and a list of locations. Vehicles have capacity fields such as weight in kilograms, volume, or units, plus travel modes (driving, truck, walking, transit) and compatibility tags. Each location needs coordinates, a time window for arrival, and a service duration in seconds. Shipment size is optional but recommended, because without it the planner may overload a vehicle.

The response lists, for each vehicle, the order sequence with estimated arrival times. It also reports metrics for each route and for all routes together: distance, time, the number and total length of delays, and cost.

Hard rules, soft rules, and one cost number

According to the documentation, hard constraints cannot be violated. They include the number of couriers, vehicle capacity, incompatibility between orders (such as special temperature needs), and a required order sequence. Soft constraints can be broken at a price: time windows, maximum mileage per shift, minimum and maximum stops, and maximum shift length. The objective is a metric named total_cost_with_penalty, where lower is better. It has two parts, the cost of resources and the cost of violations, and the service converts mileage, lateness, and route balance into one internal scale so they can be compared.

Google's OR-Tools guide on penalties and dropping visits shows the same idea in open documentation. In its 16-location example with four vehicles, a penalty of 1000 per skipped location leaves two locations unserved, and the objective of 7936 equals 5936 meters of travel plus two penalties of 1000.

How the search works

The algorithm page says the planner builds a first route, changes it by steps such as reordering orders, keeps improvements, and may also accept a worse route with a probability that decreases as the route gets worse. It repeats until it finds a good enough solution. The page does not use the name, but the pattern matches simulated annealing, which another article on this site explains. The page also states that the search is probabilistic, that identical inputs can give different routes, and that total cost varies only slightly between runs. Because checking every combination is impossible, solving time is capped and the result may not be the best one. The page adds that results are usually better than manual planning, which is the vendor's statement and not a measurement.

Why penalty weights decide the plan (illustrative)

The numbers below are invented to show the mechanism and are not RouteQ prices. A depot serves three orders. Plan X uses one vehicle and plan Y uses two. Costs: CAD 0.80 per kilometer, CAD 30 per vehicle-hour, CAD 50 fixed per vehicle used.

Plan (illustrative)DistanceVehicle-hoursLate minutesTotal at CAD 2 per late minuteTotal at CAD 12 per late minute
X: one vehicle62 km3.010CAD 209.60CAD 309.60
Y: two vehicles74 km4.00CAD 279.20CAD 279.20

Plan X costs 49.60 for distance, 90 for time, and 50 for the vehicle, which is 189.60 before lateness. With a low lateness price it wins. With a high price, plan Y wins. Neither choice is a mistake. The weights encode what the business cares about, so a user comparing services should compare the weights as well as the routes.

Limits visible in the documentation

The FAQ states that, by default, no more than 10 planning tasks can run in parallel under one access key, and it names errors for exceeding a daily or simultaneous quota without giving the quota numbers. Planning time cannot be set manually by default. The pages read for this article did not state a maximum number of orders per request. Anyone planning a large fleet would need to ask the vendor for that figure and test it.

Open tools for comparison

Several open-source engines handle the same problem family. VROOM, published on GitHub under the BSD-2-Clause license, is a C++20 engine for the traveling salesman problem, capacitated routing, time windows, and pickup-and-delivery problems, with skills, breaks, and priorities as extra rules. It can read travel times from OSRM, Openrouteservice, or Valhalla, or from a custom matrix. OR-Tools offers a routing library with the penalty mechanism shown above. A fair comparison runs the same orders through each tool and checks the plan against the same cost definition.

Questions to ask before trusting a savings claim

  • Which baseline was used: a manual plan, last month's actual routes, or a simple heuristic?
  • Were road times measured on the same data for both plans?
  • How were lateness and unserved orders priced?
  • Was the optimizer run more than once, given that the search is random?
  • Do the savings include service time at stops, which no routing tool can shorten?

FAQ

Is Yandex Routing the same as RouteQ?

A 2022 press release says RouteQ is the former Yandex.Routing, and the old documentation address redirects to RouteQ documentation. Product details may have changed since then.

Does it find the shortest possible route?

No such guarantee is documented. The search is time-capped and probabilistic, and the page itself says the best solution may not be found.

Why would the same orders give different routes on two runs?

The method uses random choices, including occasional acceptance of worse routes. The documentation says total cost differs only slightly between runs.

Can a business in Canada or the United States use it?

This article did not verify that. Terms of service, sanctions rules, and data-residency requirements should be checked with a qualified adviser before any purchase.

Last verified: October 2026

Sources

  • Routing algorithm (Russian-language page), Yandex Routing documentation, accessed October 2026
  • Vehicle routing concepts and reference, Yandex Routing documentation (English), accessed October 2026
  • RouteQ general vehicle properties and frequently asked questions, RouteQ documentation, accessed October 2026
  • RouteQ partners with NextBillion AI to increase delivery efficiency and reduce logistics costs around the world, PR Newswire, February 21, 2022
  • Penalties and Dropping Visits, OR-Tools routing guide, Google Developers, 2026
  • VROOM Project repository, GitHub, 2026

New to artificial intelligence?

Start with the basics, then follow the guides to tools, rules, and effects on society.

Start with AI Basics