3.5 times fewer qubits. That is what Quantinuum says its new architecture needs, compared with a standard surface code, to protect a logical qubit from errors. The company did not only simulate it. Quantinuum ran the demonstration on real hardware, its 98-qubit Helios processor. The result is the Quantinuum Helix error correction architecture. In the tests, encoded operations beat their unencoded physical versions without postselection. That distinction matters, because quantum error correction has spent years promising gains that show up only in simulations, or in results filtered after the fact.
The overhead problem, and Helix’s answer
Surface codes, the workhorse of quantum error correction, need many physical qubits to protect each logical qubit. That overhead is a major obstacle to building fault-tolerant machines at useful scale. Every physical qubit spent on redundancy is one that cannot run an algorithm.
Quantinuum’s engineers built a leaner design: a concatenated code they call C4-Helix. It pairs a twisted toric code with a small four-qubit block code. According to Quantum Computing Report, the design and Quantinuum’s trapped-ion shuttling give a 3.5 times lower qubit overhead than rotated surface codes of the same distance. The report gives no qubit counts that show the calculation, so the figure is a code design comparison and not something a reader can check from the published numbers.
What Helios actually showed
Quantinuum tested logical memory, logical gates and entanglement between different codes on Helios, and the authors posted the work as a preprint. The logical memory error rate came in at about 0.0046 percent per qubit per cycle, which the authors say beats the unencoded baseline (arXiv abstract). The logical Clifford gate error rate was about 0.028 percent, roughly 4.3 times better than the 0.12 percent of Helios’s physical gates (Quantum Computing Report). A three-logical-qubit GHZ state reached a fidelity with a 99.925 percent lower bound and a 99.975 percent upper bound, above the 99.537 percent physical baseline (Quantum Computing Report and the arXiv abstract).
Those error rates carry wide error bars. The arXiv abstract gives the memory error as 4.6 x 10-5 with a range of plus 6.2 and minus 2.6. The 10-6 to 10-8 logical error range that appears in coverage comes from simulations of future, better hardware, and no one has measured it.
Why It Matters to You
Nothing here is a product or a price. Lower hardware overhead could, in theory, help bring quantum computing costs down, which would matter to anyone who might someday rent a slice of a quantum computer. Picture Jordan, an illustrative example and not a real person, a logistics manager at a mid-size trucking company who has never thought about qubits but does think about the cost of cloud compute. Quantinuum has made no pricing claim, and this post can’t make one either. For now the sensible thing to watch is the qubit-overhead numbers, not just the qubit-count headlines.
Postselection, the missing pieces and the road to Apollo
Postselection means discarding the failed runs and reporting only the survivors. It has padded plenty of headline claims in this field, so reviewers now ask whether every run was kept. The authors say it was, for the headline comparisons. One figure does use it: Quantum Computing Report says a lower memory error of 1.9 x 10-5 depends on discarding 0.5 percent of runs with a decoder filter, and that number should be treated separately.
These are Quantinuum’s own results, posted as an arXiv preprint dated Sep 2, 2026, and no independent replication or peer review has turned up yet. Terms like “world-record” come from the announcement and its coverage, and they have not been checked against other labs’ results. The tests cover memory, Clifford gates and entanglement, and Quantum Computing Report says universal computation, which needs magic states, is only partly shown.
Quantinuum is not the only team pushing trapped-ion hardware this month. Germany’s JION project is bringing a microwave-controlled trapped-ion system to a German supercomputer, attacking the bottleneck from a different angle by shrinking the control hardware instead of redesigning the code. Quantum Computing Report calls the GHZ result a blueprint for Apollo, Quantinuum’s upcoming fault-tolerant hardware generation. If the 3.5 times saving holds as hardware grows past 98 qubits, error correction would become less of a bottleneck. That is a plan, not a result. Quantum Computing Report covered the full result set this week.
Quick answers
What does 3.5 times lower overhead mean?
Quantinuum says its Helix code needs about 3.5 times fewer physical qubits than a rotated surface code of the same distance to protect a logical qubit.
Was this tested on real hardware?
Yes. The tests ran on Quantinuum’s 98-qubit Helios trapped-ion processor. Only the future error-rate projections came from simulation.
Does this mean fault-tolerant quantum computers are here?
No. The authors describe an architecture for the early fault-tolerant era. Full universal computation has not been shown.
Sources
- Quantum Computing Report: Quantinuum Demonstrates Helix Quantum Error Correction Architecture on Helios Hardware (Sep 9, 2026)
- arXiv: Experimental validation of a compact fault-tolerant architecture for trapped ions (Sep 2, 2026)
Related Reading
- JION Brings Microwave-Controlled Trapped-Ion Computing to a German Supercomputer
- Diraq and Equinix Plan a Silicon Spin Quantum Computer in a Sydney Data Center
- MIT Arm Qubit Design Aims for Faster Quantum Operations, in Simulation
- Quantinuum’s own announcement of the Helios quantum computer that runs Helix
