Quantinuum said on October 19, 2026 that its Helios trapped-ion quantum computer protected a logical qubit with a fidelity of 99.9995% per error-checking cycle, equal to about five logical failures per million cycles. The company calls it a world record and a tenfold reduction from its previous reported figure. It is a memory benchmark, not a demonstration of a finished useful computation, and the study itself says so.
What was measured
Physical qubits are fragile. Error correction spreads one piece of information across many physical qubits to form a logical qubit, and extra measurements reveal signs of errors without reading out the protected data. The test is whether larger codes actually perform better. Below an error-correction threshold they should; above it, the extra operations add more noise than they remove.
Quantinuum’s researchers tested color codes at distances three, five and seven. The logical memory error rate per cycle fell from roughly 5.5 in 10,000 at distance three, to 3.8 in 100,000 at distance five, to about five in a million at distance seven. The largest memory experiment used 73 physical qubits, including measurement ancillas, to protect one logical qubit. Across the tested distances, the error rate improved by a factor of roughly 12 each time the code distance increased by two.
How to read the error bars
The headline number rests on few failure events. The estimate came from three failures across 29,000 experimental shots, each with 20 correction rounds. The paper reports a central estimate of five per million with a one-standard-deviation interval from about three to nine per million. That is strong evidence that protection is improving with code size, but it is a statistical estimate, not a guarantee. It also does not mean a million arbitrary quantum gates have run without error.
A separate experiment on two-qubit Clifford gates showed error suppression of about nine times between distances three and five. The coverage stresses these are different benchmarks, so the memory figure should not be applied to the gate result. The distance-seven gate experiment was not run because it exceeded Helios’s available qubit count.
Why Helios can try different codes
Helios uses trapped ions, whose connectivity allows interactions beyond a fixed nearest-neighbor layout. That flexibility lets researchers change error-correction strategy without redesigning the processor around a single code. The work also tackles leakage, where an ion leaves the states used to represent a qubit and can cause correlated errors that hurt correction. The team combined optical repumping, which returns leaked ions to the computational states, with circuit techniques that limit how errors spread.
The AI angle: decoding the errors
Error correction needs a classical decoder that reads the measurement patterns and infers the likely errors. The study compared several decoders, including an A* search decoder called Tesseract, a streaming decoder called Frontier, a three-dimensional convolutional neural network and a round-token transformer. The neural models were trained on data from a Helios emulator and applied to experimental data without fine-tuning on hardware data.
The reported neural inference times are in the hundreds of microseconds, which the researchers say is compatible with real-time operation on this class of hardware. An important caveat: all decoding in the study was performed offline. Streaming tests fed measurements round by round to assess timing, but the neural decoders were not controlling a live feedback loop. That remains an engineering milestone ahead. What the result shows is that neural networks trained in simulation can interpret errors from a real quantum machine.
What remains
Quantinuum says the accuracy is now within the target range for Apollo, its planned universal fault-tolerant system targeted for 2029; its roadmap also lists Sol in 2027. These are development targets. A practical machine must protect many logical qubits, run a broad set of gates, sustain correction during computation and offer enough throughput for valuable workloads. The company’s own blog identifies scalable, below-threshold magic-state preparation, needed for a route to universal fault tolerance, as work still to be demonstrated.
What to watch next
- Peer review and independent scrutiny of the failure counts behind the five-per-million estimate.
- Any live, real-time demonstration of neural decoders closing the feedback loop.
- Logical two-qubit gate results at distance seven once qubit counts allow.
- Magic-state preparation below threshold.
- Progress on the Sol (2027) and Apollo (2029) systems.
Key takeaways
- Helios logical memory fidelity reached 99.9995% per cycle, about five failures per million, a tenfold improvement on Quantinuum’s prior record.
- The estimate is based on three observed failures; the one-sigma range is roughly three to nine per million.
- Error rates improved about 12 times per two-step increase in code distance.
- Neural decoders worked on real data after training on an emulator, but decoding was offline.
- This is a memory benchmark, not a full fault-tolerant computation.
Frequently asked questions
What is a logical qubit?
A logical qubit encodes information across several physical qubits so that errors can be detected and corrected, making it more reliable than any single physical qubit.
Does 99.9995% mean quantum computers are ready for real work?
No. It shows memory protection improving with code size. Useful computing also needs many logical qubits, a full gate set and sustained correction.
Where does AI come in?
Neural-network decoders help interpret error measurements quickly. They do not replace the quantum processor.
