Changes in v9.1.1:
- Fixed attribute type for "Fingerprint" attribute in Java, C++, and .NET
- Fixed a numerical issue in a knapsack presolve reduction
- Fixed issue with IIS algorithm using too many threads
- Fixed wrong sign for Farkas dual values for variables with infinite lower bounds
- Fixed numerical issue in presolve coefficient strengthening
- Fixed issue with losing binary status of a fixed variable when writing to an mps file
- Fixed numerical issue in a dual presolve reduction
- Fixed segmentation fault when using pre-specified user cuts in some situations
- Fixed wrong answer when using pre-specified user cuts with >= sense
- Fixed handling of user cuts with equality sense
- Allow the use of a "pip install" trial license within Docker
- Fixed performance bottleneck in Compute Server model updates with general constraints
- Fixed segmentation fault when environment creation is used incorrectly by the user and the error code is not handled properly
- Fixed wrong default value of the "Lazy" attribute in attribute files
- Fixed incorrect presolve fixing for pure fixed charge network models
- Only use single-threaded BLAS on Mac to avoid overshooting the thread limit
- Fixed numerical issue for outer approximation cuts for quadratic constraints
- Fixed wrong answer for non-convex MIQCPs when sparsifying Q in objective
- Upgrade to curl 7.74.0 to address a network performance issue on Windows

Changes in v9.1.0:
- Fixed issue with empty error messages from within callback in Python
- Fixed numerical issue in cut lifting
- Added missing Visual C++ runtime dependency for Python on Windows
- Fixed issue with resetting runtime attribute to zero on model updates
- Fixed potential segmentation fault when running out of memory
- Fixed issue in gurobipy with MLinExpr objects used for objective functions of multi-objective models
- Fixed numerical issue in clique based presolve reduction that lead to wrong claim of infeasibility
- Fixed issue with cbStopOneMultiObj() not working correctly on multi-objective models solved with distributed MIP
- Fixed potential segmentation fault when calling cbStopOneMultiobj() after a SIGINT has been received
- Fixed rare performance issue in a primal heuristic that may lead to a long time without display output
- Fixed numerical issue in cut scaling that very rarely lead to not calling the callback to allow for lazy constraint generation for a solution
- Fixed performance issue in flow path cut separator for models with very special structure
- Fixed the use of feasRelax() on multi-objective models
- Fixed incorrect handling of variables with infinite lower bound and finite upper bound for QP simplex
- Fixed wrong answer bug due to a bad symmetry reduction on models with quadratic or SOS constraints
- Fixed performance issue and not checking the time limit while lifting binary variables into cuts
- Fixed issue in IIS computation where hitting the time limit on the initial solve reports an error result
- Fixed numerical issue in a knapsack based cut strengthening routine that may lead to wrong answers
- Fixed issue in gurobipy with using a singleton MVar as objective
- Fixed numerical issue in presolve implied bounds reduction
- Fixed segmentation fault due to numerical issue in crossover for handling of dual super basic variables in a singular basis
- Fixed performance issue with sub-MIP cuts taking very long on MIQPs
- Fixed issue with temporarily using more threads than allowed by the "Threads" parameter
- Fixed issue with "gurobi_cl --tokens" for single-use licenses on Windows
- Fixed issue with keeping internal bilinear constraint representation in model produced by presolve method
- Fixed issue with MIP cleanup on models with semi-continuous or semi-integer variables
- Fixed issue with getting a Q_NOT_PSD error for MIP starts applied to non-convex models
- Fixed issue with querying OBJVAL attribute for unbounded models if a feasible solution is available
- Fixed issue with returning error 10017 for an MIQCP that is presolved to an MILP and solved with distributed MIP
- Fixed wrong answer due to applying incompatible symmetry and disconnected components reductions at the same time
- Fixed rare uninitialized memory read in presolve for non-convex quadratic objective function
- Fixed issue in primal simplex ratio test that can lead to wrong conclusion of infeasibility
- Fixed issue that caused two worker processes to run simultaneously on the Compute Server for distributed MIP
- Fixed issue with getting an unbounded ray from a simplex solve
- Fixed issue with querying char attributes in gurobipy via model.getAttr()
- Fixed potential uninitialized memory read when setting MIP start via Start attribute on Compute Server

Changes in v9.0.3:
- Fixed bug in presolve with merging small cliques that can lead to a wrong answer
- Changed gurobipy argument name of addMConstrs() from "names" to "name" to match documentation
- Fixed a performance issue with an infinite loop in a heuristic due to numerical issues
- Fixed a small memory leak for multi-scenario models
- Fixed LP iteration count attribute values for multi-objective model solves
- Fixed issue in gurobipy with wrong result when adding two MLinExpr expressions
- Fixed bug with querying attributes of a multi-scenario model from Compute Server
- Fixed issue with wrong MIP solution status if a parallel thread at the root node proves infeasibility of the model
- Fixed OverflowError in gurobipy when adding many variables with addVars()
- Fixed segmentation fault when using multiple environments on Windows with a Compute Server or Instant Cloud application
- Fixed rare segmentation fault in a set partitioning heuristic
- Fixed the handling of the "ImproveStartTime" parameter for multi-objective models
- Fixed wrong answer due to bad presolve reduction for non-convex MIQCPs with quadratic equations
- Fixed segmentation fault when using user cuts in certain situations
- Fixed segmentation fault related to semi-continuous variables
- Fixed extensions in library names in the R source package on Mac
- Fixed translation of one particular form of the addGenConstrIndicator() function in gurobipy when a linear expression with a non-zero constant term is involved
- Fixed wrong answer for non-convex MIQCPs due to an invalid dual reduction on the McCormick constraints
- Fixed performance issue in gurobipy with Compute Server when printing a set of variables if no feasible solution exists
- Fixed performance issue in symmetry computation
- Fixed issue with not stopping immediately if cbStopOneMultiObj() is called from the callback for a multi-objective solve that is currently performing symmetry detection
- Fixed segmentation fault that arises from a numerical issue in presolve for MIQPs
- Fixed issue with not printing the problem size for concurrent MIP > Fixed issue with writing a solution file based on the "SolFiles" parameter if the user callback returned an error
- Added missing support of scalar division for the MLinExpr object of gurobipy
- Fixed bug with accepting a solution with big constraint violation that came from a certain heuristic
- Fixed bug with returning the scaled solution from barrier, if the barrier solve is interrupted prematurely
- Fixed segmentation fault in presolve for non-convex MIQPs
- Enable work-around for a long running model.free() call via Compute Server that would terminate the connection due to a time out
- Fixed wrong fingerprint value when writing to mps files due to handling of IEEE -0.0 (negative zero) values
- Fixed wrong answer coming from relax-and-lift cuts
- Fixed issue with duplicate variable indices in the linear constraint of an indicator constraint
- Fixed wrong answers due to invalid symmetry detection for models with SOS constraints
- Fixed segmentation fault with solving a model in batch mode that has no name
- Fixed wrong answer in non-convex MIQCPs due to a wrong local update of the McCormick constraints
- Fixed numerical issue in presolve for models with SOS constraints and big-M coefficients
- Fixed performance issue in the LP file reader when huge files are being read
- Fixed wrong answer for multi-scenario models that decompose into multiple large components