State of Quantum Hardware — Week 32, 2026

Published 2026-08-08 · auto-generated from the public corpus · subscribe

This week's fleet

The Provenova corpus holds 28 device snapshots from six hardware providers: IBM contributes 17, IQM and Origin Quantum three each, Quantinuum and IonQ two each, and Rigetti one. The most recent corpus refresh attempt is dated 2026-07-21, so nothing on the boards below post-dates that attempt. Values come from two source classes and the distinction governs how each should be read: vendor-reported figures are manufacturer claims that Provenova does not independently verify and cannot redistribute as raw data, while Metriq entries are community benchmark submissions published under CC-BY-4.0 and are redistributable.

Two-qubit gate fidelity — higher is better, every entry vendor-reported. On the two-qubit fidelity board, Quantinuum's H2 ranks first at 0.999 over 56 qubits, captured 2026-07-03, and carries a vendor-reported Quantum Volume of 33,554,432. IonQ's Forte follows at 0.996 (36 qubits, 36 algorithmic qubits) and IQM's Garnet at 0.9951 (20 qubits), both also captured 2026-07-03. Rigetti's Ankaa-3 is fourth at 0.995 over 84 qubits, but its capture date is 2024-12-23 — more than eighteen months older than every other entry on the board — so placing it in a strict ordering compares devices measured in different calibration epochs. IonQ's Aria closes the five at 0.994 (25 qubits, 25 algorithmic qubits), captured 2026-07-03. All five are manufacturer claims sourced from quantinuum.com, ionq.com, rigetti.com and the AWS Braket listing for Garnet; none is an independently verified measurement.

EPLG — error per layered gate, lower is better, every entry Metriq under CC-BY-4.0. IBM's ibm_boston leads the EPLG board at 0.00154 (156 qubits, captured 2026-02-18), ahead of Quantinuum's h2-2 at 0.00180 (56 qubits, 2026-01-23). ibm_pittsburgh records 0.00273 (156 qubits, 2026-02-19), ibm_torino 0.00551 (133 qubits, 2026-02-19), and IQM's iqm_garnet 0.00823 (20 qubits, 2026-03-02). Note that iqm_garnet and the Garnet entry on the fidelity board are separate corpus entries under different backend identifiers and different source classes; the platform does not aggregate them, and this report does not treat them as one device.

CLOPS — circuit layer operations per second, a throughput measure where higher is better, every entry Metriq under CC-BY-4.0. The CLOPS board is entirely IBM: ibm_fez at 368,905, ibm_kingston at 362,372, ibm_boston at 360,574, ibm_pittsburgh at 358,613, ibm_marrakesh at 353,642 and ibm_torino at 345,670. All are 156-qubit devices except ibm_torino at 133 qubits, and all six captures fall between 2026-02-17 and 2026-02-19.

QAOA approximation ratio — higher is better, every entry Metriq under CC-BY-4.0. The ten-entry QAOA board is led by h2-2 at 0.8018 (56 qubits), ahead of ibm_boston at 0.6734 and ibm_kingston at 0.6412. ibm_marrakesh (0.5731), ibm_pittsburgh (0.5695), ibm_torino (0.4560) and ibm_fez (0.4209) follow. The tail is the informative part: iqm_garnet sits at 0.1432, IQM's iqm_emerald at 0.0099 with a paired QFT accuracy of 0.1293 (54 qubits, captured 2025-11-19), and Origin Quantum's wukong_72 at -0.0350 with a QFT accuracy of 0.005 (72 qubits, captured 2026-02-24). The leaderboard reports the ratio without stating its normalisation, so the sign should be read only as position relative to the other entries: of the ten submissions, iqm_emerald sits within a hundredth of zero and wukong_72 is the single entry below zero. Where a Metriq submission also carries QFT accuracy, h2-2 records 0.967.

Read across all four boards, the captures span 2024-12-23 to 2026-07-03. Four of the five vendor-reported fidelity captures are the newest values anywhere at 2026-07-03, while the fifth, Ankaa-3, is the oldest capture in the corpus; every Metriq capture falls between 2025-11-19 and 2026-03-02. No board is a contemporaneous snapshot of the fleet.

Movements

The previous published edition was Week 30 (2026-07-25). No Week 31 edition was issued, so this comparison spans two ISO weeks rather than one. Measured against the figures printed in that edition, the benchmark side of the platform is static.

All four boards carry the same entries, the same ordering and the same values as Week 30, and no captured_at advanced on any of them — every capture date that edition printed is identical in this run's data, so no device re-benchmarked into or out of a ranking across the two weeks. The leaders are unchanged: H2 on vendor-reported two-qubit fidelity, ibm_boston on EPLG, ibm_fez on CLOPS and h2-2 on QAOA ratio.

The device corpus is flat as well: 28 snapshots from six providers, with the same breakdown as Week 30 (IBM 17, IQM 3, Origin Quantum 3, Quantinuum 2, IonQ 2, Rigetti 1). The most recent corpus refresh attempt is still dated 2026-07-21, the same date the Week 30 edition reported, which means no refresh has landed in the eighteen days since.

The one change is on the research side. The Week 30 text stated a total of twelve research cards; the corpus-status endpoint now reports 17, so five cards were added across the two weeks between editions, one of them inside ISO week 32.

New on the platform

One research card carries a creation date inside ISO week 32 (2026-08-03 to 2026-08-09):

The card is explicit about its own scope: it records a deterministic simulator run on Provenova inspired by that paper. It does not reproduce the paper's hardware results, and the paper's authors have not reviewed or endorsed it. What was run is a static, fully unitary 10-qubit Bernstein-Vazirani circuit — nine data qubits plus one phase-kickback ancilla — with a hidden string chosen by Provenova rather than taken from the paper. The card notes it uses 9 bits instead of the paper's 10 because the textbook n+1-qubit construction at 10 bits exceeds the platform's 10-qubit limit.

Run metadata as stated on the card: backend local_sim / aer_statevector (simulator), 4096 shots, Hellinger fidelity 1.0, verdict "reproducible", against a calibration captured 2026-01-01. Those two figures describe the simulator run measured against its own ideal reference distribution; they say nothing about the paper, and the card states directly that nothing on it measures, checks or supports the paper's hardware figures. The card adds two caveats of its own: the ancilla is never uncomputed, so in the ideal statevector the weight splits evenly across two 10-bit outcomes that agree on qubits 0 through 8; and the simulated backend carries a calibration-driven error model, so sampled shots need not land only on that ideal support.

Four further cards were created after the Week 30 edition but outside week 32. The research corpus now holds 17 cards in total, against a daily publication cap of 3.

Method note

Every figure above carries a source class, and the class governs both how much weight it bears and what may be reused. Vendor-reported specifications are manufacturer claims drawn from provider material; Provenova does not independently verify them, and they are not redistributable as raw data. Metriq entries are community benchmark submissions published under CC-BY-4.0 and are redistributable. The corpus also ingests IBM device calibration under Apache-2.0, but no value in this edition arrives that way: every IBM row on the boards above is a Metriq submission carrying CC-BY-4.0 terms.

All Provenova platform algorithm runs behind research cards, including this week's card, are deterministic simulator executions rather than hardware executions. A card's result reflects the compiled circuit and the error model applied to it, not the behaviour of a live device, and should not be read as a hardware measurement.

Capture dates are not synchronised across the corpus, and the 2024-12-23 to 2026-07-03 spread noted above means rankings compare devices measured at different points in their calibration history. Every figure in this report was fetched from the Provenova leaderboard and corpus-status endpoints on 2026-08-08 UTC.