Q-CTRL has run a functional Quantum Fourier Transform on up to 100 physical qubits in an IBM Heron R3 processor while testing a compilation method designed to limit routing and noise.
The correct frequency bitstring still emerged as the most likely measurement result after Q-CTRL ran a functional Quantum Fourier Transform on up to 100 physical qubits. The experiment used IBM's Heron R3 superconducting processor ibm_boston and applied a compilation design intended to keep a long-range algorithm within a line of locally connected qubits.
The work is identified as arXiv:2608.05435, an arXiv preprint rather than a peer-reviewed paper. It is described as the largest experimental execution of a functional QFT algorithm on quantum hardware to date, but that status should not be confused with a demonstration of fault-tolerant quantum computing or practical quantum advantage.
Its central claim is narrower than a general quantum-computing advantage: Q-CTRL demonstrated an experimental QFT subroutine that resolved periodic signal frequencies across a reported 2100-dimensional Hilbert space. The computation was performed in hardware, and the correct integer frequency remained the highest-probability output at 100 qubits, with a signal-to-noise ratio above one.
A QFT converts information about periodicity into a frequency pattern that can be read from measurements. In a real processor the circuit is constrained by the hardware graph. When two qubits cannot interact directly, compilation normally inserts additional operations to move quantum information through neighboring sites. Those routing gates increase circuit depth and create more opportunities for control errors and decoherence.
That connectivity problem is the real subject of the work. It is also the type of systems-level issue that distinguishes a physical-qubit demonstration from the logical-qubit benchmarks pursued by major research communities at MIT, CERN, and other large laboratories.
On a linear nearest-neighbor layout the proposed compilation uses identity cancellations to remove the usual routing overhead. For a full n-qubit QFT, the construction reaches the all-to-all theoretical baseline of n2 - n two-qubit CX gates. For an approximate QFT that retains the d largest phase rotations, it requires d(2n - d - 1) CX gates. Matching the gate-count baseline does not make the processor all-to-all connected; it means the compilation avoids paying an additional routing cost on the line.
Q-CTRL then modified the layout into what it calls the Convolutional AQFT. The design adds one ancilla qubit and two CX gates, reversing the direction of the initial entangling dependency chain. The resulting circuit can be treated as a translation-invariant kernel spanning d + 2 qubits. Moving that kernel along the register with a stride of one reduces the average number of two-qubit gates in each qubit's causal history.
That arrangement matters because errors do not stay confined to the gate where they begin. A shorter causal history can limit how much noise reaches later operations, although it does not eliminate gate errors or establish fault tolerance. The circuit also leaves qubits idle for substantial periods, so those windows became part of the control problem rather than an irrelevant pause.
The compilation result is therefore best interpreted as an architectural optimization. Similar distinctions are important in peer-reviewed quantum-engineering literature, including work published in journals such as Nature, where gate counts, circuit depth, noise models, and logical error rates are treated as separate performance variables rather than compressed into a single qubit-count headline.
The experiment protected idle qubits with robust dynamical-decoupling sequences designed to suppress decoherence and crosstalk. The sequences were combined with an "as late as possible" scheduling strategy so that operations were placed near the end of available timing windows. This is error mitigation and noise management, not quantum error correction: the experiment did not create logical qubits or demonstrate a fault-tolerant algorithm.
The numerical benchmarks are substantial but difficult to compare with a single headline fidelity. Across 20 randomly sampled target frequencies, the readout-mitigated process fidelity was 11.4% at 50 qubits and 1.8% at 80 qubits. The reported signal peak was 10.8 times the background at 50 qubits and 7.5 times the background at 80 qubits. At 100 qubits the target state remained the mode output, meaning it was the most frequently observed result even though the reported material does not provide a corresponding process-fidelity value.
Those figures show why output identification and state fidelity must be separated. A process fidelity of 1.8% at 80 qubits is low as a measure of agreement with the ideal process, yet the intended frequency can still stand above the background in the sampled distribution. The experiment therefore demonstrates a usable signal-extraction regime for this subroutine under the reported conditions, not a high-fidelity execution of an arbitrary 100-qubit computation.
The reported summary does not establish conventional inferential statistics such as p-values or confidence intervals, and the 20-frequency sample should not be treated as a complete characterization of every possible input. Because the result remains a preprint, independent replication and peer review will be important for assessing how robust the observed scaling is across devices, target frequencies, calibration settings, and noise conditions.
The connectivity issue also appears in a related hardware report, where processor geometry and error suppression are likewise central to interpreting a qubit-scale result. The comparison is limited: that work concerned surface-code error suppression, while this experiment concerns compilation and execution of a QFT subroutine.
The Convolutional QFT adds only two CX gates to the low-routing construction while introducing an ancilla. Its significance is consequently architectural rather than a claim that 100 physical qubits have become 100 reliable computational units. The result does not report logical-qubit encoding, logical error rates, operating temperature, coherence times, gate fidelities, or a comparison against a complete classical workflow.
Nor does the demonstration establish practical quantum advantage. The task is a controlled algorithmic benchmark on a superconducting processor, and the reported material does not provide total runtime, energy use, a competing classical implementation, or independent replication. Those omissions matter because a compact circuit can be scientifically valuable without delivering a faster or cheaper solution to an external problem.
Q-CTRL's broader product strategy also emphasizes physics-informed AI for autonomous quantum-computer calibration, including its Boulder Opal platform for tuning and maintaining processors. That commercial context helps explain the focus on calibration, scheduling, compilation, and hardware performance, but it does not change the evidentiary status of this particular QFT result.
The strongest evidence is that compilation, scheduling, and dynamical decoupling can work together to preserve the dominant output of a large QFT circuit despite restricted connectivity and accumulating noise. The engineering limitation is equally clear: the process fidelity falls sharply between 50 and 80 qubits, while the 100-qubit claim rests on identifying the correct mode rather than reproducing the full quantum process with high fidelity.
A physical qubit is one hardware element that can be controlled and measured; a logical qubit stores information across multiple physical qubits to detect or correct errors. This experiment used physical qubits and did not report logical qubits or an error-correcting code. Its convolutional strategy is therefore best understood as a way to reduce exposure to noise before correction becomes available. That is a meaningful compilation result, but it is not evidence that a fault-tolerant quantum computer has been built.
Q-CTRL's demonstration earns attention because it attacks a concrete bottleneck rather than presenting raw qubit count as capability. It shows that a carefully structured QFT can preserve a recognizable frequency signal on a restricted-connectivity processor at 100 physical qubits, while the low process fidelities and missing classical comparison keep the claim firmly in the laboratory-demonstration category. The practical lesson is direct: better compilation can extend what noisy hardware reveals, but it cannot substitute for reliable logical qubits, reproducible performance, or a demonstrated advantage on a useful task.
