Mittelmann’s optimization benchmarks
The independent solver benchmarks maintained by Hans Mittelmann (Arizona State University) — for decades the reference for how fast optimization solvers actually are. T2: an academic third party publishing comparative results, not a vendor.
Categories run well past LP and MILP: combinatorial optimization (TSP, CVRP), semidefinite and SOCP, nonlinear programming, mixed-integer quadratic and quadratically-constrained problems, mixed-integer nonlinear, and problems with equilibrium constraints. Results are aggregated by shifted geometric mean rather than performance profiles, with time limits and hardware stated per report, and the interface can show a “virtual best (ensemble)” — what you would get by always picking the right solver.
The finding that is not about performance
The page records its own partial dismantling by the vendors it measures. “Through an action by Gurobi at the 2018 INFORMS Annual Meeting this has come to an end. IBM and FICO demanded that results for their solvers be removed.” Gurobi and MindOpt withdrew in 2024; other developers have filled some of the gap.
So the field’s longest-running independent comparison can no longer publish results for several of the leading commercial solvers, at the vendors’ insistence. This spoke should file that alongside its own benchmark scepticism: no-free-lunch-theorem says comparisons must be made against a stated problem class, and population-optimization-benchmark shows how much a ranking depends on the suite — but here the constraint is not methodological at all. The comparison is incomplete because the strongest participants exercised a licence term.
It also puts the vendor-published speed claims that circulate in this field in their place: they exist partly because the independent alternative was made unavailable. Pair with miplib, which supplies the instances.
Current figures pulled 2026-08-10 (quality cycle)
Two reports read in full, so the spoke now holds numbers and not just a description of the page:
- LPopt (2026-07-01, 65 disclosed instances + 16 undisclosed): COPT 1.0 (65 solved) · MOSEK 6.24 · XOPT 6.63 · HiGHS 12.8 (54) · CLP 27.4 · GLOP 59.7 · SoPlex 104 (31).
- MILP (2026-07-07, 240 MIPLIB2017 instances, 2 h on a Ryzen 9 5900X): COPT 1.0 (219) · Optverse 1.72 (210) · XSMOO 5.15 · HiGHS-parallel 5.44 (179) · SCIPCO 6.59 · HiGHS 7.55 (158) · SCIP 10 at 9.93 (136).
The index is actively maintained — LPfeas and Large-SOCP both carry 2026-08-10 dates, i.e. the day this was read.
The vendor-withdrawal caveat bites here. COPT is the baseline in both tables, so every ”×” figure above is relative to the fastest solver that still allows publication. With Gurobi, CPLEX, Xpress and MindOpt absent, the true commercial ceiling is unmeasurable from this source, and the open-source gap recorded on highs is a lower bound.