HiGHS
The open-source LP/MIP/QP solver from the University of Edinburgh, MIT-licensed, C++11. Until this page the spoke had a metaheuristics corpus with no solver in it — linear-programming and integer-programming rested on Wikipedia, and nothing here could actually be run. T1 (the project's own documentation).
What it is
Four separate solvers under one library, each with a named author:
| Component | Method | Author |
|---|---|---|
| LP | dual revised simplex | Qi Huangfu |
| LP | interior point | Lukas Schork |
| QP | active set | Michael Feldmeier |
| MIP | branch-and-cut | Leona Gottwald |
Julian Hall runs the project; Ivet Galabova maintains the engineering. Built for large-scale sparse
models, serial and parallel. Interfaces from C, C#, Fortran, Julia and Python, plus a standalone
executable on Windows, Linux and macOS. That interface list is why HiGHS is the default solver behind
SciPy's linprog and much of the Python optimization stack — most people running it have never
typed its name.
The site claims "high performance" and publishes no numbers of its own, which is the ordinary vendor-documentation problem. Mittelmann supplies them.
Where it actually places (Mittelmann, 2026)
LPopt (optimal basic solution, 65 disclosed instances, 2026-07-01) — scaled shifted geometric mean, COPT as 1.0:
| Solver | Ratio | Solved /65 |
|---|---|---|
| COPT (commercial) | 1.0 | 65 |
| MOSEK (commercial) | 6.24 | 52 |
| XOPT | 6.63 | 52 |
| HiGHS 1.15.0 | 12.8 | 54 |
| CLP | 27.4 | 40 |
| GLOP (OR-Tools) | 59.7 | 33 |
| SoPlex | 104 | 31 |
MILP (240 MIPLIB2017 instances, 2 h limit, Ryzen 9 5900X 12-core, 2026-07-07):
| Solver | Scaled GM | Solved /240 |
|---|---|---|
| COPT | 1.0 | 219 |
| Optverse | 1.72 | 210 |
| XSMOO (FiberSCIP hybrid) | 5.15 | 174 |
| HiGHS 1.15.1 parallel | 5.44 | 179 |
| SCIPCO | 6.59 | 153 |
| HiGHS 1.15.0 | 7.55 | 158 |
| SCIP 10 | 9.93 | 136 |
Two readings. First, HiGHS is the strongest fully open solver in both tables — it beats CLP by 2×, GLOP by 4.7× and SoPlex by 8× on LP, and beats SCIP on MILP while solving 22 more instances. Second, the gap to the commercial front is not small: 12.8× on LP, 7.55× on MILP, and the free solver leaves 82 of 240 MILP instances unsolved inside two hours that COPT finishes. Open-source has won the usable default position, not the fastest one.
Read the ranking with mittelmann-benchmarks' own caveat attached: Gurobi, CPLEX and Xpress do not appear in these tables because their vendors demanded removal. The commercial baseline here is COPT — the fastest solver that still permits publication, which is not necessarily the fastest solver.
Why it matters to this spoke
The corpus is almost entirely metaheuristics — GA, PSO, simulated-annealing and the long tail of nature-inspired variants, all of which return good answers with no optimality guarantee. HiGHS is the other tradition: for problems that fit the LP/MIP form, an exact method with a proof, benchmarked on a public library (miplib) at a scale the metaheuristics literature never attempts. The no-free-lunch-theorem applies across problem classes; within the class where simplex and branch-and-cut apply, structure beats search and it isn't close.
The engine paper (2026-09-03)
huangfu-hall-parallel-simplex is the peer-reviewed account of the dual revised simplex code this solver is built on, published in Mathematical Programming Computation in 2018. It supplies what a project page cannot: the prior consensus that the revised method was unsuitable for parallelisation, the two schemes (PAMI across iterations, SIP within one) that got around it, and measured factors on 30 mostly-Mittelmann problems rather than a claim.
Related
linear-programming · integer-programming · mittelmann-benchmarks · miplib · huangfu-hall-parallel-simplex · convex-optimization · metaheuristic-optimization · no-free-lunch-theorem · synthesis