What is Quantinuum’s Helios and what is measured?

TL;DR (status as of October 16, 2026): Helios is Quantinuum's commercial trapped-ion quantum computer: 98 fully connected barium-137 qubits whose measured two-qubit gate fidelity of 99.921% was peer-reviewed in Nature 655 (June 17, 2026). It launched commercially on November 5, 2025 (quantinuum.com). What is measured: component fidelities, 94 error-detected and 48 error-corrected logical qubits — with more than 96% of shots discarded in the most complex runs (arXiv:2602.22211) — and a certified-randomness amplification protocol on 64 qubits (arXiv:2511.03686). What is promised: Apollo, a universal fault-tolerant machine, by 2030 (Quantinuum press release, September 10, 2024). What no external judge has yet published a verdict on: any of it — Quantinuum remains in Stage B of DARPA's Quantum Benchmarking Initiative, outside the six-team final stage announced October 7, 2026 (darpa.mil). Rosetta Quantum keeps the three columns — measured, promised, judged — separate.

What is Helios?

Helios is the machine Quantinuum sells today: a quantum charge-coupled device (QCCD) trapped-ion processor with 98 barium-137 hyperfine qubits, ytterbium-171 ions as sympathetic coolant, and all-to-all connectivity through physical ion transport (Nature 655, 81–86, 2026). It was launched commercially in New York on November 5, 2025, and is available through Quantinuum's cloud service and as an on-premises offering (Quantinuum press release, November 5, 2025). The vendor's product page lists a base-unit power draw below 40 kW without support infrastructure (quantinuum.com, accessed October 16, 2026) — for scale, our survey of quantum power draw measured ≈26 kW for a full dilution-refrigerator system of a different architecture. Research customers named at launch: Amgen, BMW Group, JPMorganChase and SoftBank Corp. (press release, November 5, 2025).

The launch framing deserves its own sentence. Quantinuum presented Helios as enabling «Generative Quantum AI», stating that «Helios unlocks the ability to enhance GenAI models with quantum generated data» (press release, November 5, 2025). That sentence is written in the present tense; the evidence is not there yet. As of October 1, 2026, our verdict on quantum computing and generative AI counted zero measured quantum speedups on generative-AI workloads in the public record. Whether quantum-generated data will improve GenAI models is a question that will be settled by measurements that do not exist today.

What is measured on Helios?

The strongest numbers are component fidelities, and they are peer-reviewed. The Nature paper (vol. 655, pages 81–86, published June 17, 2026) reports a single-qubit gate infidelity of 2.5(1)×10⁻⁵, a two-qubit gate infidelity of 7.9(2)×10⁻⁴ — that is the 99.921% in our hero — and a state-preparation-and-measurement (SPAM) infidelity of 3.3(5)×10⁻⁴, all averaged over the machine's operational zones (Nature 655, 2026). The same paper is candid about scope: only the perfect-entangler RZZ(π/2) gate was benchmarked, and the authors write that «we do not yet fully understand the power or limitations of Helios» (Nature 655, 2026).

Two more measured numbers put the fidelities in context. First, speed: the average depth-1 circuit time is about 55 ms per layer (Nature 655, 2026), and a separate experiment on the same machine measured about 0.913 s from first gate layer to bitstring on a 64-qubit, two-qubit-depth-9 circuit (arXiv:2511.03686). Ion traps buy fidelity at the price of wall-clock speed — the trade our superconducting vs trapped-ion comparison mapped. Second, the classical frontier: the paper reports random-circuit-sampling results and estimates, via tensor-network contraction cost under stated assumptions about memory and GPU efficiency, that Helios «operates well beyond the reach of classical simulation» (Nature 655, 2026). That estimate of the classical rival is self-reported, and the history of such estimates is erosion: Google's 2019 figure of 10,000 years fell to days, then to parity. We print the claim with its author and its assumptions attached.

What do «94 logical qubits» actually mean?

They mean something specific and narrower than the headline. The launch-day logical-qubit results run on iceberg codes — a family that encodes k logical qubits in k+2 physical ions (arXiv:2602.22211). The distance-2 versions can only detect errors, not correct them: when an error is detected, the shot is discarded and restarted. The 94 «error-detected» logical qubits were globally entangled in a GHZ state with an acceptance rate of about 0.25 — three out of four shots thrown away (arXiv:2602.22211, via postquantum.com, March 10, 2026). The 48 «error-corrected» logical qubits come from a concatenated [[80,48,4]] code at distance 4, with a QEC-cycle acceptance of about 0.62; the deepest XY-model circuits accepted about 3.2% of shots, and the analysis states that «more than 96% of the computational shots are discarded» in the most complex experiments (postquantum.com, March 10, 2026). The same analysis is explicit on what this is not: not fully fault-tolerant computing, with physical gate errors around 10⁻⁴ against the ≈10⁻⁸–10⁻¹⁰ it cites as needed for practical fault tolerance.

The more recent result is stronger on quality and smaller in count. In September 2026, Quantinuum demonstrated its Helix architecture on Helios: a [[20,2,6]] C4-Helix code — 2 logical qubits in 20 ions — running the complete two-qubit logical Clifford group, with a logical two-qubit Clifford error of 2.8×10⁻⁴ versus 1.2×10⁻³ physical, a ≈4.3× improvement (arXiv:2609.03194, via quantumcomputingreport.com, September 9, 2026). Non-Clifford gates and magic-state distillation were not demonstrated; the coverage ties those to the future Apollo generation (quantumcomputingreport.com, September 9, 2026). The count-versus-quality tension is exactly why the physical-to-logical ratio, not the headline number, decides the timing — and why logical-qubit demos from different vendors are not directly comparable.

What has Helios run beyond its own benchmarks?

One applied protocol with named external users: certified randomness amplification. A JPMorganChase and Quantinuum team used Helios as the remote quantum server in a protocol that amplifies a weak randomness source into nearly perfect randomness — 64 qubits, 276 two-qubit gates, an XEB score of 0.586, measurement bases revealed about 30 ms before readout, and a min-entropy rate of β = 0.53 against a restricted adversary (arXiv:2511.03686, November 6, 2025). The authors print their own fine print: «the ability to perform beyond-classical sampling does not, by itself, enable certified randomness», the security rests on computational assumptions about classical simulation cost, and «a complexity-theoretic proof remains an open challenge» (arXiv:2511.03686). This is the amplification echo of the only applied quantum-advantage claim still standing in Rosetta's registry — and it inherits that claim's asterisks.

What is promised, and in which tense?

The promise has a name and a date. On September 10, 2024, Quantinuum published its accelerated roadmap «to achieve universal, fully fault-tolerant quantum computing by 2030», with a fifth-generation system, Apollo, that will be «fully fault-tolerant and universal» and will execute «circuits with millions of gates»; the roadmap, per the company, «draws a direct line to hundreds of logical qubits» (Quantinuum press release, September 10, 2024). Those are commitments, and they will stay in the promised column of our scoreboard until hardware measurements move them. The product page we accessed on October 16, 2026 names no dated intermediate system between Helios and that goal (quantinuum.com) — the 2030 promise lives in the 2024 press release.

Where does an external judge place Quantinuum today?

In the waiting room, not the final. Quantinuum entered Stage B of DARPA's Quantum Benchmarking Initiative in the November 2025 cohort. On October 7, 2026, DARPA moved Atom Computing, Diraq, IBM and IonQ into Stage C — joining Microsoft and PsiQuantum for a six-team final stage — and Quantinuum was not among them (darpa.mil, October 7, 2026; postquantum.com). Two fairness notes come with that fact: DARPA advances companies in a staggered way, so absence from one announcement is not a rejection (postquantum.com, October 2026), and the QBI's verification results are not public in any case, as our explainer on the program documented. The net position, stated plainly: the strongest Helios numbers are vendor-measured and peer-reviewed, which is the top of the self-reported range on the 2026 hardware map — and no external judge has yet published a verdict on any Helios capability claim.

The spec sheet, with the status of each number:

What we don't know

• Whether the fidelities hold off the benchmark. Only the RZZ(π/2) gate was benchmarked, averaged over zones; the Nature authors state they do not yet fully understand the machine's power or limitations (Nature 655, 2026). A Publisher Correction was issued on July 7, 2026, and the page we accessed does not state what it corrects.
• Whether the «beyond classical simulation» estimate survives. It depends on stated assumptions about memory and GPU efficiency (Nature 655, 2026), and no adversarial classical team has yet published a counter-estimate for Helios that we know of. The 2019 precedent took months to erode.
• A Quantum Volume figure. We did not find a Helios Quantum Volume value in the sources we accessed; the product page links to a separate data page we could not read in this run (quantinuum.com, accessed October 16, 2026).
• Whether Quantinuum will advance to QBI Stage C. DARPA's process is staggered and more promotions are expected (postquantum.com, October 2026); a Stage B seat today says nothing final.
• What Apollo will ship. «Millions of gates» and «hundreds of logical qubits» are roadmap language from 2024; the measurements that would move them to the teal column will exist only when the hardware does.

Sources

• Quantinuum et al. — «A 98-qubit trapped-ion quantum computer with all-to-all connectivity», Nature 655, 81–86 (published 2026-06-17) — nature.com
• Quantinuum — «Quantinuum Announces Commercial Launch of New Helios Quantum Computer…» (2025-11-05) — quantinuum.com
• Quantinuum — Helios product page: 98 qubits, fidelities, power, access (accessed 2026-10-16) — quantinuum.com
• Quantinuum — «Quantinuum Unveils Accelerated Roadmap to Achieve Universal, Fully Fault-Tolerant Quantum Computing by 2030» (2024-09-10) — quantinuum.com
• Liu et al. (JPMorganChase / Quantinuum) — «Certified randomness amplification by dynamically probing remote random quantum states», arXiv:2511.03686 (2025-11-06) — arxiv.org
• PostQuantum — «Quantinuum Squeezes 94 Logical Qubits from 98 Physical — But What Does It Actually Mean?», analysis of arXiv:2602.22211 (2026-03-10) — postquantum.com
• Quantum Computing Report — «Quantinuum Demonstrates Helix Quantum Error Correction Architecture on Helios Hardware», coverage of arXiv:2609.03194 (2026-09-09) — quantumcomputingreport.com
• DARPA — «Four more teams enter Quantum Benchmarking Initiative final stage» (2026-10-07) — darpa.mil
• PostQuantum — «DARPA Quantum Benchmarking Stage C» — Stage B/C rosters and the staggered-advancement note (October 2026) — postquantum.com
• Rosetta Quantum — related ledger entries: #74 the surviving advantage claim · #80 DARPA QBI · #62 quantum and generative AI · #33 logical qubits and the ratio · #13 vendor metrics · #55 the 2019 supremacy claim · #50 the 2026 hardware map · #56 power draw · #20 superconducting vs trapped ion

Rosetta Quantum publishes verdicts with raw, reproducible data. This is educational content, not a product claim. Quantum, independently measured.