#linear-programming #optimization #symbolic #solver


Mixed Integer Linear Programming for Rust, with an user-friendly API. This crate allows modeling LP problems, and lets you solve them with various solvers.

48 releases (18 stable)

1.7.0 Oct 6, 2023
1.5.0 Aug 25, 2023
1.4.1 Jul 18, 2023
1.4.0 Mar 12, 2023
0.4.3 Mar 19, 2021

#13 in Math

Download history 7475/week @ 2023-08-12 7578/week @ 2023-08-19 8642/week @ 2023-08-26 8917/week @ 2023-09-02 9251/week @ 2023-09-09 8325/week @ 2023-09-16 5799/week @ 2023-09-23 8548/week @ 2023-09-30 9104/week @ 2023-10-07 9164/week @ 2023-10-14 10704/week @ 2023-10-21 11842/week @ 2023-10-28 15262/week @ 2023-11-04 14298/week @ 2023-11-11 12437/week @ 2023-11-18 9365/week @ 2023-11-25

52,121 downloads per month
Used in 30 crates (4 directly)

MIT license



A Mixed Integer Linear Programming modeler that is easy to use, performant with large problems, and well-typed.

Crates.io documentation MIT license

use std::error::Error;

use good_lp::{constraint, default_solver, Solution, SolverModel, variables};

fn main() -> Result<(), Box<dyn Error>> {
    variables! {
               a <= 1;
          2 <= b <= 4;
    } // variables can also be added dynamically
    let solution = vars.maximise(10 * (a - b / 5) - b)
        .using(default_solver) // multiple solvers available
        .with(constraint!(a + 2 <= b))
        .with(constraint!(1 + a >= 4 - b))
    println!("a={}   b={}", solution.value(a), solution.value(b));
    println!("a + b = {}", solution.eval(a + b));

Features and limitations

  • Linear programming. This crate currently supports only the definition of linear programs. You cannot use it with quadratic functions. For instance: you can maximise 3 * x + y, but not 3 * x * y.
  • Continuous and integer variables. good_lp itself supports mixed integer-linear programming (MILP), but not all underlying solvers support integer variables. (see also variable types)
  • Not a solver. This crate uses other rust crates to provide the solvers. There is no solving algorithm in good_lp itself. If you have an issue with a solver, report it to the solver directly. See below for the list of supported solvers.


Pull requests are welcome ! If you need a feature that is not yet implemented, get in touch. Also, do not hesitate to open issues to discuss the implementation.


If you need non-linear programming, you can use lp-modeler. However, it is currently very slow with large problems.

You can also directly use the underlying solver libraries, such as coin_cbc or minilp if you don't need a way to express your objective function and constraints using an idiomatic rust syntax.

Usage examples

You can find a resource allocation problem example in resource_allocation_problem.rs.


This library offers an abstraction over multiple solvers. By default, it uses cbc, but you can also activate other solvers using cargo features.

solver feature name integer variables no C compiler* no additional libs** fast
highs ✅+
  • * no C compiler: builds with only cargo, without requiring you to install a C compiler
  • ** no additional libs: works without additional libraries at runtime, all the dependencies are statically linked
  • + highs itself is statically linked and does not require manual installation. However, on some systems, you may have to install dependencies of highs itself.

To use an alternative solver, put the following in your Cargo.toml:

good_lp = { version = "*", features = ["your solver feature name"], default-features = false }


Used by default, performant, but requires to have the cbc C library headers available on the build machine, and the cbc dynamic library available on any machine where you want to run your program.

In ubuntu, you can install it with:

sudo apt-get install coinor-cbc coinor-libcbc-dev

In MacOS, using homebrew :

brew install cbc

Be careful if you disable the default features of this crate and activate the cbc feature manually. In this case, you have to also activate singlethread-cbc, unless you compiled Cbc yourself with the CBC_THREAD_SAFE option. Otherwise, using Cbc from multiple threads would be unsafe.


minilp is a pure rust solver, which means it works out of the box without installing anything else.

Minilp is written in pure rust, so you can use it without having to install a C compiler on your machine, or having to install any external library, but it is slower than other solvers.

It performs very poorly when compiled in debug mode, so be sure to compile your code in --release mode when solving large problems.


HiGHS is a free (MIT) parallel mixed integer linear programming solver written in C++. It is able to fully leverage all the available processor cores to solve a problem.

good_lp uses the highs crate to call HiGHS. You will need a C compiler, but you shouldn't have to install any additional library on linux (it depends only on the C++ standard library). More information in the highs-sys crate.


lp_solve is a free (LGPL) linear (integer) programming solver written in C and based on the revised simplex method.

good_lp uses the lpsolve crate to call lpsolve. You will need a C compiler, but you won't have to install any additional library.


The lp-solvers feature is particular: it doesn't contain any solver. Instead, it calls other solvers at runtime. It writes the given problem to a .lp file, and launches an external solver command (such as gurobi, cplex, cbc, or glpk) to solve it.

There is some overhead associated to this method: it can take a few hundred milliseconds to write the problem to a file, launch the external solver, wait for it to finish, and then parse its solution. If you are not solving a few large problems but many small ones (in a web server, for instance), then this method may not be appropriate.

Additionally, the end user of your program will have to install the desired solver on his own.


SCIP is currently one of the fastest open-source solvers for mixed integer programming (MIP) and mixed integer nonlinear programming (MINLP). It is also a framework for constraint integer programming and branch-cut-and-price. It allows for total control of the solution process and the access of detailed information down to the guts of the solver.

good_lp uses SCIP through the its rust interface russcip. To use this feature you will need to install SCIP. The easiest way to do it is to install a precompiled package from here or through conda by running

conda install --channel conda-forge scip

Variable types

good_lp internally represents all variable values and coefficients as f64. It lets you express constraints using either f64 or i32 (in the latter case, the integer will be losslessly converted to a floating point number). The solution's values are f64 as well.

For instance:

// Correct use of f64 and i32 for Variable struct and constraints
  variables! {
      a <= 10.0;
      2 <= b <= 4;
  let model = problem
    .with(constraint!(a + 2 <= b))
    .with(constraint!(1 + a >= 4.0 - b));

Here, a and b are Variable instances that can take either continuous (floating-point) or integer values. Constraints can be expressed using either f64 or i32, as shown in the example (but replacing for example 4.0 with a usize variable would fail, because an usize cannot be converted to an f64 losslessly).

Solution values will always be f64, regardless of whether the variables were defined with f64 or i32. So, even if you use integer variables, the solution object will store the integer variable values as f64.

For example, when printing the solution:

// Correct use of f64 for solution values
println!("a={}   b={}", solution.value(a), solution.value(b));
println!("a + b = {}", solution.eval(a + b));

// Incorrect use of i32 in combination with solution value (Will fail!)
println!("a + 1 = {}", solution.value(a) + 1); // This will cause a compilation error!

The solution.value(a) and solution.value(b) will return f64 values, and solution.eval(a + b) will also provide an f64 value.


This library is published under the MIT license. The solver themselves have various licenses, please refer to their individual documentation.


~109K SLoC