- We've added support for Gurobi and HiGHS. 
- The implementation is a simple prototype, which we will use as our starting 
point. 
- We ran some experiments comparing Gurobi, HiGHS, CPLEX, and Soplex on the 
default optimal benchmark suite. 
- A branch with the prototype is here:  
https://github.com/remochristen/downward/tree/lpsolvers_hack

When we changed the LP interface a few years back from OSI to our custom interface, the performance of SoPlex dropped by 13 tasks. We have been trying to analyze this in issue1124 and tracked it down to changes and bugs in SoPlex. Since you are running tests with SoPlex versions 6 and 8, it would be good to also check if we regain the lost performance with this upgrade. Of course more things could have changed between version 6 and 8 but at least the bug fixed in version 6.0.4 should also be fixed in version 8.

We have some initial results to share, which can be found here:

https://ai.dmi.unibas.ch/_experiments/gustavo/sprint2026_lpsolvers/data/sprint20
26_lpsolvers-benchmark_solvers-eval/sprint2026_lpsolvers-
benchmark_solvers.html#summary

A complete description of the experimental setup and an analysis of the results 
are in the attached file.

We want to explore LP solvers, such as Gurobi, HiGHS, Google OR-Tools, and FICO 
Xpress. So far, we have implemented interfaces for Gurobi and HiGHS. The 
implementations currently use some 'hacks' to change two-sided constraints into 
one-sided versions during model creation. As one of the next steps, we want to 
change the LP interface to only support one-sided constraints.