State of Quantum Hardware — Week 33, 2026

This week's fleet

The corpus behind these boards holds 28 device snapshots across six providers: IBM 17, Origin 3, IQM 3, Quantinuum 2, IonQ 2, Rigetti 1. IBM supplies well over half the corpus, so several boards below are largely an IBM-vs-IBM comparison rather than a cross-vendor one.

The status endpoint reports last_refresh_attempt as 2026-07-21T23:09 UTC, against a server time of 2026-08-15T23:11 UTC. No refresh attempt has been recorded in roughly the last three and a half weeks, and the field records an attempt rather than a confirmed success. Treat everything below as a static snapshot of what was already in the corpus, not a reading of this week's hardware.

Two-qubit fidelity (higher is better) is vendor-reported throughout — manufacturer claims that Provenova has not independently verified, redistributable_raw: false, with the licence reference pointing at the vendor's own site for Quantinuum, IonQ and Rigetti, and at aws.amazon.com/braket for IQM Garnet. Quantinuum H2 leads at 0.999 on 56 qubits, with a reported Quantum Volume of 33,554,432, captured 2026-07-03. IonQ Forte follows at 0.996 on 36 qubits (36 algorithmic qubits), then IQM Garnet at 0.9951 on 20 qubits, both also captured 2026-07-03. Rigetti Ankaa-3 sits fourth at 0.995 on 84 qubits — but captured 2024-12-23, about eighteen months before its board-mates. Ranking it against figures from mid-2026 is not a like-for-like comparison, and the 0.0001 gap separating third from fourth place is far smaller than the uncertainty a capture that old introduces. IonQ Aria closes the board at 0.994 on 25 qubits (25 algorithmic qubits), captured 2026-07-03.

The other three boards are metriq throughout — CC-BY-4.0 · metriq.info, redistributable_raw: true: community benchmark submissions, redistributable, but filed by third parties under conditions the board does not fully describe.

EPLG (error per layer — lower is better) is led by ibm_boston at 0.00154 on 156 qubits (captured 2026-02-18), ahead of h2-2 at 0.00180 on 56 qubits (2026-01-23) and ibm_pittsburgh at 0.00273 on 156 qubits (2026-02-19). ibm_torino records 0.00551 on 133 qubits (2026-02-19) and iqm_garnet 0.00823 on 20 qubits (2026-03-02).

CLOPS (higher is better) is six IBM systems and nothing else, so it says nothing about cross-vendor throughput. ibm_fez tops it at 368,905 (156 qubits, 2026-02-17), followed by 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. Top to bottom the whole board spans about 6.7% — narrow enough that the ordering may not survive a recalibration.

QAOA approximation ratio (higher is better) is led by h2-2 at 0.802, then ibm_boston at 0.673 and ibm_kingston at 0.641, with ibm_marrakesh at 0.573, ibm_pittsburgh at 0.570 and ibm_torino at 0.456. ibm_fez — first on CLOPS — places seventh at 0.421: circuit throughput does not predict solution quality. iqm_garnet is eighth at 0.143.

The tail is where the caveats bite. iqm_emerald records 0.00992 on 54 qubits, with a QFT accuracy of 0.129, captured 2025-11-19 — the oldest capture on the board. wukong_72 returns −0.0350 on 72 qubits (2026-02-24), alongside a QFT accuracy of 0.005. The endpoint does not state how the ratio is normalised, so a negative value should be read only as a position on this board — wukong_72 is the single entry below zero, and iqm_emerald the only other within a hundredth of it — not as a performance figure, and not as evidence about the hardware. A near-zero or negative entry is a reason to go read the underlying submission, not to rank it.

Finally, the same physical machines appear twice under different identifiers from different sources: Quantinuum as vendor-reported H2 and metriq h2-2, and IQM as vendor-reported Garnet and metriq iqm_garnet. The platform stores these as separate entries and does not aggregate them; readers should not merge them either.

Movements

Week 32 was published 2026-08-08 and this run's data was fetched 2026-08-15, so the interval is a single week. Against the figures that edition printed, one thing changed.

All four boards are unchanged: the same entries, in the same order, at the same values, with the same capture dates wherever Week 32 printed one, and with the same leaders — H2 on vendor-reported two-qubit fidelity, ibm_boston on EPLG, ibm_fez on CLOPS, h2-2 on QAOA ratio. No captured_at advanced anywhere, so nothing re-benchmarked into or out of a ranking this week.

The device corpus is flat too: 28 snapshots, with the same per-provider breakdown Week 32 reported. The most recent corpus refresh attempt still carries the 2026-07-21 stamp that edition printed — and Week 32 in turn recorded it as unchanged from Week 30 — so the refresh date has not advanced across three consecutive editions.

The single change is the research card total, which went from the 17 Week 32 reported to 18. The added card was created 2026-08-09, the day after Week 32 published and still inside ISO week 32.

New on the platform

As of the corpus-status fetch at 2026-08-15T23:11 UTC, no research card had been published in ISO week 33 (2026-08-10 to 2026-08-16). The endpoint reports zero cards created on 2026-08-15, and none of the 18 cards it lists carries a created_at timestamp inside that range. The platform total stands at 18 published cards, against a daily publishing cap of 3.

One card did land after the Week 32 edition went out on 2026-08-08, so readers of that edition will not have seen it. It belongs to ISO week 32, not this one: p=1 QAOA MaxCut ansatz on a 4-cycle, created 2026-08-09T23:34 UTC and dated "Published 2026-08-09" on its page.

The card references arXiv:2608.06212, "Warm-Starting MaxCut Relaxation via Low-Depth Quantum Approximate Optimization Algorithm" by Bao G. Bach, Ilya Safro and Filip B. Maciejewski (2026). It states its own scope plainly: a deterministic, fixed-seed simulator run inspired by that paper, not a reproduction of it. The run used the local_sim / aer_statevector backend at 4,096 shots, recording a Hellinger fidelity of 1.0 and a verdict of reproducible.

What it does not do is worth stating in full. It does not run the paper's Erdős–Rényi or Sherrington–Kirkpatrick instances, nor its n = 500 and n = 1000 scales; the depth-one ansatz and the four-vertex cycle are Provenova's own choices. The card prepares an ansatz state only and does not compute the local correlators the paper's method extracts. No hardware is involved. The paper's authors were not involved in the card and have not reviewed or endorsed it.

Method note

Three source classes sit behind the figures above, and the class determines both how much weight a number carries and whether the raw value may be reused. Metriq entries are community benchmark submissions published under CC-BY-4.0 and are redistributable. Vendor-reported specifications are manufacturer claims taken from provider material; Provenova does not independently verify them, and they are not redistributable as raw data. The corpus also ingests IBM device calibration under Apache-2.0.

In this run the split falls cleanly along board lines. The EPLG (5 entries), CLOPS (6) and QAOA-ratio (10) boards are Metriq submissions throughout. The 2Q gate fidelity board (5 entries) is vendor-reported throughout, so every value on it is an unverified manufacturer claim. No entry on any of the four boards arrived as IBM calibration data this week — the IBM rows are all Metriq submissions.

Research cards are deterministic simulator executions, not hardware runs, and say nothing about a live device.

The boards are not a contemporaneous snapshot: capture dates in this run span 2024-12-23 to 2026-07-03. All figures were fetched 2026-08-15 UTC from the leaderboard and the corpus-status endpoint.