Brion Q.A.I.™
Quantum Artificial Intelligence. A trained quantum neural network — run it yourself.
Live on Brion Statevector Simulator
Run Q.A.I.-1
Four measurements go in as rotation angles on four qubits. The circuit runs on Brion Statevector Simulator for 2048 shots, and the parity of the measured bits is the model's output. Nothing is cached — every press executes the circuit again, which is why two runs of the same input differ slightly.
Result
Press run. The measurement histogram below is the raw output of the circuit — the prediction is derived from it, not the other way round.
—
Measurement counts — 2048 shots
Model card — Q.A.I.-1
Every figure here is read from the deployed artifact, which was written by the training run itself. Accuracy is quoted next to what it has to beat.
| Task | iris-versicolor-vs-virginica |
|---|---|
| Data | sklearn.datasets.load_iris (Fisher 1936, UCI) |
| Classes | versicolor, virginica |
| Held-out accuracy | 96.7% on 30 samples |
| Majority-class baseline | 50.0% |
| Logistic regression, same split | 96.7% |
| Qubits × layers | 4 × 3 |
|---|---|
| Trained parameters | 24 |
| Readout | register_parity_z |
| Optimiser | SPSA |
| Circuits run in training | 56,000 |
| Trained on | brion-sim |
| Trained at | 2026-08-19T20:48:06+00:00 |
| Commit | 13018a7 |
What this actually is
A variational quantum classifier. The four input measurements are written
into the angles of RY rotations, one per qubit. A trained
circuit of 24 rotation parameters and a ring of
CNOT gates then acts on that state, and every qubit is measured. The
model's output is the parity of the measured bits, averaged over
2048 shots — a single number between −1 and +1. Above the
fitted threshold means one class, below means the other.
The CNOT ring is the part that makes this a quantum model rather than four independent rotations. Without it the circuit could not represent any correlation between the four features, and the parity readout would carry no more information than the features separately.
What it is not
It is not faster than a classical model at this task, and it does not beat one: logistic regression scores 96.7% on the same split, which is the same as the quantum model, on a laptop, instantly. Four qubits is a state space of sixteen amplitudes; a classical computer holds that without noticing. That number is on the model card because a benchmark quoted without its classical reference is not a result.
What it does demonstrate is a complete pipeline: parameters fitted by running real circuits, an artifact carrying its own provenance, and a served endpoint whose answer comes from a circuit executed at request time. That pipeline is the thing that scales to a problem where the quantum side might matter — the model on it today is a small one.
Run Q.A.I.-1 from the API
curl -X POST https://brionquantum.com/api/v1/models/qnn-hybrid/inference \
-H "X-API-Key: $BRION_API_KEY" \
-H "Content-Type: application/json" \
-d '{"input": [6.3, 3.3, 6.0, 2.5]}'Verified on quantum hardware
A GHZ state was prepared at a range of widths on Origin Wukong, a 166-qubit superconducting processor, at 2,048 shots per circuit on 2026-08-24. Every row below carries the provider's own job ID, so any of these runs can be looked up rather than taken on trust.
| Qubits | GHZ population | Machine | Provider job ID |
|---|---|---|---|
| 2 | 0.9990 | WK_C180 | CDF796D005747808EA701C0301917847 |
| 4 | 0.9946 | WK_C180 | 0617BFA8049217C0EB13EC93D1925FCC |
| 8 | 0.8550 | WK_C180 | 8D300259316371EBCD46EA71A3B32ABF |
| 12 | 0.0000 | WK_C180 | B409E172787AC8B013E7BCA371C7A043 |
| 16 | 0.0005 | WK_C180 | A6844171053F78E6D0DB321DD771B406 |
| 24 | 0.0000 | WK_C180 | FAED9F313B39123D6B37EB227FECA5D5 |
| 32 | 0.0000 | WK_C180 | 9AAB2C88229DB99E1F0ACCCF0E404EEB |
| 48 | — | — | Origin returned neither counts nor probabilities for the job: query qcloud task... |
| 64 | — | — | Origin returned neither counts nor probabilities for the job: query qcloud task... |
| 96 | — | — | Origin returned neither counts nor probabilities for the job: query qcloud task... |
| 128 | — | — | Origin returned neither counts nor probabilities for the job: query qcloud task... |
| 166 | — | — | Origin returned neither counts nor probabilities for the job: query qcloud task... |
Read it honestly, because the failures are the informative part. Entanglement survives to 8 qubits and then collapses: at 12 qubits and beyond the GHZ population is zero, and the widest requests were refused by the provider outright. The device advertises 166 qubits; what it will hold in one coherent entangled register, on this run, was 8. That gap is the real state of the hardware, and no amount of qubit count on a spec sheet closes it.
The metric is also narrower than it looks. GHZ population is P(all-zero) + P(all-one) — an upper bound on fidelity, not fidelity itself, and it does not test coherence. These are raw device counts with no readout mitigation applied.
One more thing worth stating plainly: the deployed model above was trained on the Brion statevector simulator, not on this processor — 56,000 circuits under SPSA at 512 shots each. The hardware run on this page verifies the fleet and the submission path, not the fitting of those particular parameters.
What Q.A.I. means
Most work filed under “quantum machine learning” keeps the model classical and reaches for a quantum computer to accelerate one step inside it — a sampler, a kernel, a linear solve. The network still lives in floating-point arrays on a GPU; the QPU is a subroutine it calls.
Q.A.I. is the other arrangement. The model is the circuit. Its parameters are rotation angles, its forward pass is a circuit execution on quantum hardware, and its output is a measurement statistic. There is no classical network underneath being helped along. The model above is served that way today: every press of Run compiles the input into rotation angles and executes the circuit on Brion Statevector Simulator, which is why two identical inputs return slightly different numbers.
Being precise about the current state, because the distinction only means something if it is measured: the deployed model is small — four qubits, a task a laptop solves instantly — and the parts around the circuit are still classical, as feature scaling and thresholding are on any quantum classifier. What is quantum today is the inference itself, and the training that fitted it: every parameter was set by running real circuits and reading real measurement counts. Scaling that arrangement to a problem where the quantum side carries an advantage is open research, and it is the work.
The names
Q.A.I. — also written QAI — is the name Brion and the Brion Q.A.I. Lab give to their family of quantum artificial intelligence models and to the platform that serves them. What the name marks is whose models these are: everything released under it comes from this lab.
Everything released under it is served the way the model above is: a model card written by the training run itself, an accuracy quoted next to the classical baseline it has to beat, and a circuit executed at request time rather than a cached answer. That standard is what the name is meant to stand for, and it is the part worth defending.
- Mark
- Brion Q.A.I.™ · QAI™
- Owner
- Brion Quantum · Brion Q.A.I. Lab
- First public use
- , on this page
- Status
- Unregistered common-law mark (™), in use in commerce
- Goods and services
- Quantum machine-learning models and the hosted API that serves them
- First model
- Q.A.I.-1, served on the API as
qnn-hybrid
A.Q.I. — the same technology, named from the other side
A.Q.I. (Artificial Quantum Intelligence) is the second name for the same technology, and it is not a synonym. The two terms describe Brion’s work from opposite directions, and each carries something the other leaves out.
Quantum Artificial Intelligence reads the noun as artificial intelligence, and says the substrate is quantum: this is AI, and the AI is done quantum-mechanically. Artificial Quantum Intelligence reads the noun as quantum intelligence, and says the origin is engineered: this is a quantum system built to be intelligent, rather than a classical one given a quantum accelerator.
Both readings are true here, and that is the point of carrying both names. Because the model is the circuit — parameters are rotation angles, the forward pass is a circuit execution — there is no classical network underneath to be the “real” model. Q.A.I. alone would let a reader assume the familiar arrangement, an AI that calls a quantum computer. A.Q.I. closes that reading off. The pair states the architecture more precisely than either does alone.
- Mark
- Brion A.Q.I.™ · AQI™
- Owner
- Brion Quantum · Brion Q.A.I. Lab
- First public use
- , on this page
- Status
- Unregistered common-law mark (™), in use in commerce
- Goods and services
- Quantum artificial intelligence models, the agent software that runs them on quantum hardware, and the hosted API serving both
- Applies to
- The same platform as Q.A.I., named from the other direction
The agent layer is part of what both names cover. It holds the tools that execute on real hardware, the backend selection that keeps a four-qubit circuit off a GPU across the country, the readout mitigation applied to raw device counts, and a feasibility assessor that answers a question the fleet cannot do with the factor it is short by and the classical method that does solve it — rather than a hedge.
Q.A.I., A.Q.I., Brion Q.A.I., Brion A.Q.I. and Brion Q.A.I. Lab are trademarks of Brion Quantum. Other names and marks appearing on this site belong to their respective owners.