Public benchmarks
Head-to-head numbers on the standard combinatorial-optimization suites. Every run uses the same solver budget and dataset methodology. We update the published results on every release.
Benchmark pipeline online — first run scheduled.
We're populating the public leaderboard against Gurobi, D-Wave Leap Hybrid, and NVIDIA cuOpt right now. The published table will include the exact solver budget, active adapters, and per-instance outcomes once the first run is complete.
Adapters active in last run: nerox-hybrid
Methodology
- NEROX: hybrid GPU annealing on Modal. Tier auto-selected by VRAM headroom.
- Gurobi: commercial MIP via gurobipy. Same wall-clock budget as NEROX.
- D-Wave Leap Hybrid: LeapHybridSampler with the same time-limit.
- NVIDIA cuOpt: routing problems only (TSP/CVRP) — cuOpt does not solve QUBO/MaxCut natively.
- Suites: Gset (MaxCut), TSPLIB (TSP), MQLib (BQP/QUBO).
- Gap = |solver_energy − optimal| / |optimal|.
- Each release records the active adapters, wall-clock budget, and per-instance outcomes used for comparison.