MIPLIB 2017

The standard instance library for mixed-integer programming, maintained at the Zuse Institute Berlin — current maintainers Ambros Gleixner and Mark Turner. T1, first-party.

Two sets: a benchmark set of 240 instances chosen to be solvable by current solvers and numerically stable, and a much larger collection set assembled for diversity without those restrictions. Instances carry a status:

• easy — solvable within one hour on standard hardware by an out-of-the-box solver,
• hard — solved only with extended runs, special hardware or specialised algorithms,
• open — not yet reported solved by anyone.

Solution files and status updates ship in periodic releases; the site was last updated January 2026.

Why the “open” category is the interesting one

This spoke’s benchmark culture (coco-bbob, test-functions-for-optimization, population-optimization-benchmark) is built on synthetic functions with known optima, where the question is how close a search gets. MIPLIB is the opposite: real submitted instances, and a standing list of ones nobody has solved. Difficulty is not designed in, it is discovered — and the library’s three-way status is a public record of where the frontier actually is, refreshed as solvers improve.

That gives the corpus something it lacks for its metaheuristic half: a benchmark whose hardness is empirical rather than constructed, and a curated set that vendors and researchers compete on in public. The complement — running solvers against it — is mittelmann-benchmarks.