Quantum Computing Report

Aqarios Launches Constraint-Native Quantum Optimizer in the IBM Qiskit Functions Catalog

Aqarios GmbH, a Munich-based quantum software spin-off of LMU Munich, has released its Constrained Quantum Optimizer in the IBM Qiskit Functions Catalog, giving IBM Quantum Premium, Flex, and on-premises users a managed solver for discrete combinatorial optimization. Built on the company’s FlexQAOA engine and validated in New Journal of Physics, the function ingests standard LP, MPS, or Luna Model files and enforces hard one-hot and set-packing constraints natively at the circuit level rather than relying solely on quadratic penalty terms.

Aqarios Constrained Quantum Optimizer Hardware Benchmarks (IBM Heron Processors)
Benchmark WorkloadHardware & Problem ScaleMeasured Solution Quality & Runtime
Hardware-Tailored Spin Glass144-qubit ibm_boston (Heron r3); 48 one-hot constraintsFound exact optima in 3 of 5 instances and >99% approximation ratio in the remaining 2.
QOBLIB MIS (es60fst02)156-qubit ibm_marrakesh (Heron r2); 186 vars (124 qubits)100.0% feasible optimum in 147.0 s QPU time (940.2 s total runtime; 25,000 shots).
QOBLIB MIS (sloane_2dc_128)156-qubit ibm_marrakesh (Heron r2); 128 vars (74 qubits)100.0% feasible optimum in 106.0 s QPU time (347.6 s total runtime; 12,500 shots).
Karate Club Max-4-Cut156-qubit ibm_marrakesh (Heron r2); 132 vars (132 qubits)100.0% feasible optimum in 200.0 s QPU time (258.3 s total runtime; 25,000 shots).

Conventional quantum optimization workflows translate constrained problems into unconstrained binary models by adding penalty weights that expand the effective search space and require manual tuning. Aqarios’ solver instead implements Iterative Warm-Start XY-Mixer QAOA (IWS-QAOA), which aligns a biased initial state with a warm-started XY-mixer Hamiltonian to confine state evolution to the feasible Hamming-weight-1 subspace while using fixed linear-ramp schedules to eliminate variational parameter training loops on quantum hardware.

Prior to circuit synthesis, an automated preprocessor performs graph clique merging and pendant-node variable fixing to compress qubit counts, converts remaining constraints using structure-derived penalty weights, and applies greedy steepest-descent post-processing to repair noise-induced constraint violations. In Qiskit Functions Catalog evaluations on IBM Heron hardware, replacing variational training with iterative warm-starting reduced QPU compute times from 130 to 200 seconds down to roughly 30 seconds on medium and large problem tiers.

Read Aqarios’ release announcement here, view the IBM Qiskit Functions Catalog entry here, review the technical documentation here and Maximum Independent Set tutorial here, and access the peer-reviewed study in New Journal of Physics here.

November 17, 2026

Exit mobile version