You can buy QUBO solves from single vendors, who list their solvers on the AWS and Azure marketplaces, or from D-Wave, which sells access to its annealers through Leap. On Quip Network, you post a problem with a reward attached, and miners compete to send in answers, which the chain checks before paying out.
QUBO and Ising problems
A QUBO (quadratic unconstrained binary optimization) problem gives each variable a value of 0 or 1. You add up a cost from single variables and from pairs of variables, and a solver looks for the assignment with the lowest total. An Ising model describes the same problem with spins of minus one or plus one in place of 0 and 1, and a simple substitution converts one form into the other.
Where QUBO solves are sold
Each row below comes from the seller's listing or launch coverage, so the last column reports what the seller claims.
| Option | Solver behind it | Purchase model | Answer checking |
|---|---|---|---|
| BEIT QUBO Solver, AWS Marketplace | Exact GPU branch-and-bound (BEIT) | Pay per instance solved, $0.30 listed, cancel any time | BEIT states it returns the proven minimum |
| Quplane, AWS Marketplace | A suite of 97 solvers that Quplane says runs in your own AWS account | Free preview edition | Quplane attaches optimality certificates to some QUBO answers |
| Toshiba SQBM+, Azure Marketplace | Quantum-inspired simulated bifurcation on standard computers | Azure virtual machine plans | Approximate solutions, no certificate described |
| D-Wave Leap | Advantage and Advantage2 annealers, plus hybrid solvers | Free trial by application, no prices on the page | No certificate described |
| Quip Network job mempool | Competing miners on CPU, GPU and quantum annealing hardware | You set a reward per job, paid to the winning miners | The chain recomputes the energy of every answer and enforces your quality floors |
With the first four options, you pay a single vendor to run its solver. On Quip, miners with different hardware attempt the same job, and the chain pays the answers ranked best under the payout rule you chose when posting.
A worked example, from QUBO to the chain
The submit-a-job guide builds a small binary model with a bias of minus one on variable 0 and a coupling of 2 between variables 0 and 1. As a QUBO, the cost reads:
cost = -x0 + 2·x0·x1
If you try all four assignments of x0 and x1 by hand, the lowest cost comes out to minus one, at x0 = 1 and x1 = 0.
Miners solve Ising models, so the tools convert binary models to spins before submitting. After you substitute x = (1 + s) / 2 for each variable, the energy becomes:
energy = 0.5·s1 + 0.5·s0·s1
The bias on spin 0 cancels to zero, and so does the constant offset. The lowest energy is still minus one, now at s0 = +1 and s1 = -1, which maps back to x0 = 1 and x1 = 0.
For the last step, Quip encodes coefficients and energies as whole numbers of thousandths. The job carries a bias of 500 on spin 1 and a coupling of 500 between spins 0 and 1, and the optimal answer's energy travels as -1000. According to the guide, fixed-point integers let every machine on the network compute the same energy for the same answer.
Posting the job
After you install the XQuad tools with pip install "xquad[quip]", you can run a model on a free local CPU solver first. When you swap in SolverQuip, the same solve() call sends the model to the Aglais test network. The guide lists the options for a job order:
- Reward. The chain reserves it from your account when you post, holds it until the job resolves, then pays it to the winners.
- Access. Any registered solver may answer by default, or you can limit the job to named miners or to the hardware types miners declare.
- Payout rule. The single best answer takes the full reward, or the top answers split it equally or by rank.
- Quality floors. You can require a minimum number of answers per submission, an energy ceiling and a diversity floor, so a miner cannot pad a set with copies of the same answer.
Each job has a hard deadline counted in blocks. Once the first valid answer arrives, the chain opens a shorter window for other miners to beat it. Unless you use native mode, the tools place your variables onto the chain's default registered topology, so every coupling must land on one of its edges, as Ising Topologies explains.
To try it, follow Submit your first compute job. A job on Aglais costs the minimum reward of 1 AGLS plus a small fee, and the faucet hands out AGLS for free, so your first job costs no real money.
