Quantum Error Correction Benchmark Puts Quantinuum Helios on Top

An outside benchmark favors trapped ions

Most of the quantum industry's claims about error correction are graded by the companies making them. That is why a new comparison from Germany's Jülich Supercomputing Centre is worth attention. J. A. Montañez-Barrera and Kristel Michielsen tested 10 processors from Quantinuum, IBM and IQM. Michielsen also holds an appointment at the University of Cologne.1 Quantinuum's machines handled the operations behind error correction more reliably than the others, and its newest system, Helios-1, did best.18

The preprint is titled "Evaluating the performance of QEC primitives on quantum processors at large width and depth" and is listed as arXiv:2610.05928.8 It does not measure finished error correction. It measures the pieces that error correction is built from: mid-circuit measurement, qubit reset, feed-forward control (choosing the next operation based on a measurement result) and the idling of qubits that are waiting their turn.8 Those details are often missing from the qubit-count headlines that dominate coverage of quantum computing companies, and the study's conclusions depend on them.

What the Jülich team measured

The main question was how much extra error comes from measuring qubits partway through a computation. Error correction can't work without these measurements, because they are how a machine spots signs of damage. But each measurement can also disturb nearby qubits or leave other qubits sitting idle.5 To isolate that cost, the researchers ran the same task twice: once with the measurements and the actions they trigger, and once without them.1

The two vendors came out far apart. On IBM's processors, the versions with measurements had effective error rates about ten times higher than the versions without.1 On Quantinuum's machines, the extra error from measurement was about the same size as the error from an ordinary two-qubit gate.12 The study gives Quantinuum's H2-1 and Helios-1 effective error rates of roughly 0.22% to 0.49%, against about 1.1% to 3.2% for IBM's Heron and Eagle chips. IBM's newer Heron hardware did improve on its predecessors.8 IQM's Emerald and Garnet chips lost signal quickly once chains grew past two data qubits, which the study attributes to limits on conditional control and on routing.8

On a surface-code layout with 25 data qubits, Helios-1 scored higher than the comparison hardware at every circuit depth tested. The margins were two to four standard deviations per data point, with 50 runs per point.1 The team then pushed Helios-1 much further: 81 data qubits in surface-code patterns, 91 in triangular color-code patterns, and 48 in a bivariate-bicycle pattern.1 The bivariate-bicycle tests matter because that family of codes aims to cut the number of qubits error correction needs. Even at 91 data qubits, where each layer required 45 mid-circuit measurements across 98 physical qubits, the color-code test kept a usable signal.8

The benchmark itself is a modified version of the quantum approximate optimization algorithm (QAOA). Its problems are set up so that solving them requires the same connections and measurements that particular error-correcting codes use.5 Scores run from 0.5, which is no better than random, to 1, which is optimal.3

The caveats

This result should not be overstated. All the coverage agrees that the large tests showed Helios-1 can carry out demanding patterns of error-correction operations. None of them showed working error correction at those sizes.12 The 30-qubit chain comparison was also weaker than the headlines imply. Helios-1 had the lower estimated error rate, but the uncertainty was too wide for that measurement to prove an improvement by itself.1

The study also turned up a scaling problem. As the qubit chains on Helios-1 grew longer, the extra cost of measurement rose. The authors tie this to how many operations the machine can run at once.1 Helios has eight operating zones, and only four of them perform two-qubit gates.516 In the team's first runs, many operations happened one after another. When the researchers rewrote the schedule to run more gates and measurements in parallel, the benchmark showed a clear improvement.12 This is the finding I think matters most. Trapped ions are known to be accurate, and the open question is whether they can keep that accuracy as machines grow. This study measures that tension directly.

The reports also disagree on one detail: how many runs the benchmark needs. One summary says as few as 500 to 1,000 executions per test were enough to track logical-memory performance on IBM hardware.3 Other accounts stress the 50 runs per point used in the surface-code comparison.1 The two figures probably describe different experiments rather than contradicting each other. Either way, the benchmark is cheap to run compared with full error-correction experiments.

Checking the benchmark against real error correction

A new benchmark is only useful if it predicts something real. The Jülich team checked theirs on IBM's Phoenix processor. They compared benchmark scores with actual error-corrected memory experiments in 11 regions of the chip, using the same qubits and connections in the same sessions.1 Regions that scored well on the benchmark generally had fewer logical-memory errors.1 The reported rank correlation was about 0.83 to 0.88.8

On average, the benchmark placed each chip region 1.6 positions away from where the memory experiment ranked it. Standard calibration metrics missed by 2.1 to 2.7 positions.1 For another stored quantum state the advantage was smaller, and a simple measure of the worst two-qubit gate error did almost as well.15 My reading is that the benchmark looks like a good screening tool for choosing which parts of a chip to use and for testing schedules. It is not yet a replacement for running full error-correction experiments.

How it fits Quantinuum's 2026 run

The study adds to a busy year for the company. In March, Quantinuum researchers reported up to 94 error-detected logical qubits and 48 error-corrected logical qubits on the 98-qubit Helios. Logical gate error rates were about one in ten thousand, which is lower than the hardware's physical gate error.415 In June, Nature published joint work with Microsoft showing error reductions of 11 to 800 times compared with running the same circuits on unprotected physical qubits.67 In September came Helix, an error-correction architecture that reported an error of 4.6 × 10⁻⁵ per logical qubit per correction cycle, without discarding bad runs.1620

The business side changed too. Quantinuum raised $1.68 billion in a June Nasdaq IPO under the ticker QNT. It guides to only $28 million to $32 million in 2026 revenue.16 Its roadmap calls for Sol, with hundreds of qubits, in 2027 and Apollo, a fully fault-tolerant machine, in 2029.16 IBM is aiming at the same year with Starling, which it says will offer 200 logical qubits.16

The Jülich paper matters because it comes from outside the company. Independent observers have pointed out that most Helios figures come from Quantinuum's own papers, and that logical-qubit claims should be treated as vendor numbers until someone else reproduces them.1216 Here, outside researchers ran the tests. One leading trade outlet still published the summary under a press-release label, so Quantinuum is clearly using the result in its marketing.1

It is also not a full picture of the field. Google, QuEra and IonQ hardware wasn't tested. Industry trackers credit QuEra with the largest verified logical-qubit count, 96, while Quantinuum leads on encoding efficiency.1422 IonQ has separately reported beating a superconducting qLDPC experiment by up to nine times on a 40-ion chain.23 Some figures don't match across reports either. Helios's two-qubit fidelity is usually given as 99.921%, but one guide lists 99.97%.1626 Trackers also differ on whether Quantinuum's 48 logical qubits count as error-detected or error-corrected.1922

What it means for post-quantum cryptography

For security teams, the issue is overhead. One analysis argues that a 2:1 ratio of physical to logical qubits would mean breaking RSA-2048 needs only about 2,800 trapped-ion qubits, compared with millions for surface-code superconducting designs.24 That argument depends on low-overhead codes keeping their performance as machines scale up. The Jülich findings speak to that question. The color-code and bivariate-bicycle patterns held up well on Helios-1, but measurement costs rose with chain length.18 The same analyst still puts a cryptographically relevant quantum computer 7 to 15 years away, depending on hardware type.23 Because of how slowly cryptographic migrations happen, that range supports starting post-quantum work now rather than waiting.

The study shows that Quantinuum's hardware currently handles the mid-circuit work of error correction better than IBM's or IQM's. It also gives the industry a cheap, independent way to check vendor claims. The scheduling results show that Quantinuum's lead depends on how much the machine can do in parallel, and the next test is whether Sol can increase that in 2027.