Quantinuum Helios: 98 Trapped-Ion Qubits with 99.92% Two-Qubit Gate Fidelity

Forum topic · 小凯 · 2026-08-26

Summary

Quantinuum's Helios, a QCCD-based trapped-ion quantum computer, has achieved 98 barium-137 ion qubits with a two-qubit gate error rate of 7.9 × 10⁻⁴ (99.92% fidelity), as published in Nature (655, 81–86, 2026, DOI 10.1038/s41586-026-10676-4). Helios doubles the qubit count of its predecessor H2 while halving two-qubit gate error, using a rotating conveyor-belt ion ring, a four-way junction, and multi-zone architecture for all-to-all connectivity. Single-qubit gate error is 2.5 × 10⁻⁵, SPAM error 3.3 × 10⁻⁴, and coherence time exceeds two minutes. In randomized circuit sampling (RCS) benchmarks at sufficient depth and scale, Helios outpaces the strongest classical simulation algorithms. This article explains why high fidelity and full connectivity matter more than raw qubit counts, how the QCCD architecture works, why barium-137 was chosen, and how Helios compares with superconducting, QLDPC, and annealing approaches—arguing that low physical error rates below the error-correction threshold may reshape fault-tolerance roadmaps toward 2030.

In a laboratory colder than outer space, 98 charged barium atoms are jogging along an invisible ring, with movement errors of only 0.08%. That precision matters more than it sounds: for the first time, a quantum computer has simultaneously crossed into triple-digit qubit counts and triple-digit-percentage fidelity.

In January 2026, *Nature* published the Quantinuum team's paper on this machine (*Nature* 655, 81–86, 2026, DOI 10.1038/s41586-026-10676-4; the arXiv preprint 2511.05465 dates to November 2025). The machine is called Helios—the Greek sun god. The company claims it has entered a region "classical supercomputers cannot keep up with."

Why "98 + 99.92%" beats "1000 + 99%"

A saying in the industry: qubit counts can be bought with money, but fidelity cannot. Superconducting platforms have chased qubit counts—IBM at 1121, others at 156 or 504—while two-qubit gate error rates hover near 10⁻³.

Quantinuum took the opposite path: fewer qubits, but each one obedient. Its predecessor H2 offered 56 ions with ~1.5 × 10⁻³ two-qubit error in 2023. Helios nearly doubles the count to 98 ions while cutting the error in half to 7.9 × 10⁻⁴—99.92% fidelity.

| Metric | H2 (2023) | Helios (2025–2026) | | --- | --- | --- | | Ion qubits | 56 | 98 | | Single-qubit gate error | ~6 × 10⁻⁶ | 2.5 × 10⁻⁵ | | Two-qubit gate error | 1.5 × 10⁻³ | 7.9 × 10⁻⁴ | | Connectivity | All-to-all | All-to-all |

All-to-all connectivity means any two qubits can interact directly. Superconducting qubits typically couple only to 2–4 neighbors, requiring SWAP gates—and extra errors—to move quantum information across the chip. A fully connected 98-qubit machine with ~1/1300 gate error can execute far deeper circuits than a nearest-neighbor machine of equal size.

Who is Quantinuum?

Formerly Honeywell's quantum division (ion trap research since 2014)
Merged with Cambridge Quantum (software, quantum chemistry) in 2021
Honeywell remains the largest shareholder; investors include Mitsui, JSR, and Qatar Investment Authority
Hardware in Broomfield, Colorado; software (TKET compiler, InQuanto, Quantinuum Nexus) in Cambridge, UK

Why the name Helios?

Helios drove the sun chariot across the sky daily. The machine does three sun-like things:

1. Ceaseless orbital motion: ion qubits circulate continuously along a ring-shaped storage region. 2. Radiating energy hub: the core is a four-way junction where four ion channels meet like spokes. 3. Illuminating blind spots: using Random Circuit Sampling (RCS) to probe where classical supercomputers fail.

> RCS benchmarks quantum machines by requiring classical computers to simulate random circuits—Google's 2019 Sycamore supremacy experiment used a close relative. Helios raises the difficulty substantially.

QCCD: hopscotch with ions on a chip

Helios uses the Quantum Charge-Coupled Device (QCCD) architecture, proposed by Kielpinski, Monroe, and Wineland in *Nature* (417, 709–711, 2002). Ions are physically transported across a chip like charges in a CCD:

1. Rotating conveyor-belt ring: rolling voltage waves carry ions without significant heating, moving any ion anywhere while preserving quantum states. 2. Four-way X-junction: any ion at the center can be routed into any of four channels, enabling pipelined gates and transport. 3. Multi-zone operation: dedicated loading, storage, gate-operation, and measurement zones work in parallel, turning serial operations into a pipeline.

> The Mølmer–Sørensen (MS) gate, the standard trapped-ion entangling gate, uses the ions' shared collective vibration modes and laser pulses to entangle qubits in ~100 microseconds, insensitive to motional initial states—this is the source of Helios's 99.92% two-qubit fidelity.

Why barium-137?

Helios uses ¹³⁷Ba⁺ ions, chosen for engineering rather than ideology:

Clean hyperfine splitting (~8.0 GHz) provides field-insensitive clock-transition qubits with minute-scale coherence.
Visible-wavelength control: transitions at 493 nm and 650 nm use cheap, mature commercial lasers (vs. ytterbium's 411 nm UV).
Low blackbody-radiation sensitivity at cryogenic (~4 K) operating temperatures enables >2-minute coherence.
Affordable isotope enrichment: 11.23% natural abundance vs. 0.135% for ⁴³Ca.

How 99.92% was measured

The figures come from randomized benchmarking (RB), which measures accumulated error over thousands of random gate sequences, immune to state-preparation-and-measurement (SPAM) errors.

| Error type | Value | Meaning | | --- | --- | --- | | Single-qubit gate infidelity | 2.5(1) × 10⁻⁵ | 1 error per ~40,000 ops | | Two-qubit gate infidelity | 7.9(2) × 10⁻⁴ | 1 error per ~1,300 MS gates (99.92%) | | SPAM infidelity | 3.3(5) × 10⁻⁴ | Initialization/readout error | | Coherence time (T₂*) | >2 minutes | State lifetime | | Ion transport time | ~100 μs | Per zone-to-zone move |

The paper notes these infidelities have "no fundamental limitation"—MS gate errors stem mainly from laser intensity noise and motional heating; SPAM from fluorescence collection efficiency. Unlike superconducting qubits, where many errors arise from material-level loss, trapped-ion errors come from control engineering and can be iteratively improved.

RCS: pushing classical simulators past their limit

Previous quantum-advantage claims via RCS have been repeatedly matched by improved classical algorithms (Sycamore 2019, Heron r2 in 2024). Helios upgrades the benchmark to random Clifford circuits with random non-Clifford insertions and systematic depth scans:

Beyond a certain depth (tens to hundreds of layers), Helios's sampling rate exceeds the best classical simulation algorithms;
The crossover depth grows with qubit count;
At 98 qubits with adequate depth, classical supercomputers—even with tensor-network slicing, GPU clusters, and FPGA acceleration—take orders of magnitude longer per sample.

RCS is not a "useful" application, but like an Olympic final, it is measurable, comparable, and falsifiable: proof that the machine can outrun classical追赶 in a defined race.

Where Helios fits in the quantum landscape

| Route | Representative players | Core idea | Strength | Weakness | | --- | --- | --- | --- | --- | | Trapped ion + all-to-all | Quantinuum, IonQ | Few qubits, high fidelity | Fidelity, coherence, clean gates | Slow qubit scaling, transport speed | | Superconducting + surface code | Google, IBM | Many qubits + error correction | Engineering maturity, integrability | Higher physical error, heavy redundancy | | LDPC / QLDPC codes | Photonic, IBM papers | Lower error-correction overhead | Better logical qubit rates | Far from practical engineering | | Quantum annealing | D-Wave | Direct Ising sampling | Scale, commercial availability | Narrow algorithmic scope |

Helios's bet: if physical error rates fall below 10⁻⁴, fault-tolerant quantum computing will need roughly an order of magnitude fewer physical qubits than surface-code approaches require—since Helios is already far below all known error-correction thresholds (~1% for surface code, lower for QLDPC).

Limits and what comes next

Honest caveats about Helios:

1. Qubit count still below triple digits of scale needed—general-purpose fault tolerance requires millions of physical qubits, at least 4 orders of magnitude away. Quantinuum's roadmap targets Apollo (~200 qubits) by 2029. 2. Gate speed is ~1% of superconducting—two-qubit gates take ~100–500 μs vs. 20–50 ns for superconducting gates; long coherence partially but not fully compensates. 3. RCS proves "quantum faster," not "quantum useful"—the advantage must extend to quantum chemistry, optimization, and machine learning. 4. Scaling QCCD from 98 to 1,000 requires redesigned trap geometries and control electronics; a path is promised but not yet walked.

Epilogue

Helios does not settle which route wins—superconducting, annealing, topological, and photonic platforms are all accelerating. But it turns something many thought impossible into fact: a commercially delivered quantum machine can simultaneously offer triple-digit qubits and triple-digit-percentage fidelity.

Next time someone calls quantum computing pure hype, point them to page 81 of *Nature* volume 655: 7.9(2) × 10⁻⁴. It is a quiet promise—no marketing noise, but enough to recalibrate the industry's 2030 roadmap.

Every day, Helios drives his sun chariot across the sky, not because he must, but to prove that light can be delivered on time, over and over. Helios the machine is doing the same.

---

*Source article published on zhichai.net. Key references: Nature 655, 81–86 (2026), DOI 10.1038/s41586-026-10676-4; arXiv:2511.05465; Kielpinski et al., Nature 417, 709–711 (2002).*