On October 19, 2026, Quantinuum announced that an experiment on its Helios quantum computer achieved a logical memory fidelity of 99.9995%. The figure corresponds to a logical error rate of roughly 5 in 1 million per error-checking round, and the experiment confirmed on real hardware that errors decrease as the protection from multiple physical qubits is strengthened. However, this does not mean the company ran 1 million general-purpose quantum operations in a row without failure. The number measures the reliability of a memory that holds quantum information. To judge how far this result moves practical quantum computing forward, it is necessary to distinguish how the error rate was calculated from the technical challenges that remain in operations and control.
The 99.9995% fidelity was calculated from 3 failures in 29,000 trials
The largest logical memory experiment used 73 physical qubits, including auxiliary qubits for measurement. These were combined to form a single logical qubit that protects quantum information from errors. Helios uses barium-137 ions trapped by electric fields as its physical qubits.
In the experiment, the team prepared the logical qubit in six states, positive and negative, corresponding to the X, Y, and Z measurement bases. They then repeated "syndrome extraction," which detects signs of errors, for 20 rounds, and finally measured whether the quantum information had been preserved correctly. This procedure was carried out 29,000 times in total. Analysis with a decoder called "Tesseract," the main one used, found 3 logical failures in the final readout. A decoder is a mechanism that analyzes the error signs obtained from measurements and estimates the errors that occurred in the quantum information.
Because each of the 29,000 trials involved 20 rounds of checks, the total number of check rounds was 580,000. The team also counted errors associated with preparing the initial state as half a round, and used the approximation "number of failures ÷ number of trials ÷ 20.5" to calculate a logical error rate of about 5×10⁻⁶ per round. Converting this into fidelity gives the announced figure of 99.9995%.
However, this error rate is a statistical estimate and carries some uncertainty. The Wilson confidence interval equivalent to one standard deviation (about 68%) given in the paper runs from about 3×10⁻⁶ to 9×10⁻⁶. Converted to failures per 1 million rounds, the central value is about 5, while the interval corresponds to roughly 3 to 9. Because only 3 failures were actually observed, the 99.9995% figure should not be taken as precise down to its last digit. Adam Reed, a physicist at Quantinuum, also explained on the company's official blog that when errors become this rare, each individual observed failure becomes extremely important.
Details of the research are compiled in a public manuscript by Quantinuum's Ali Lavasani, Adam P. Reed, and colleagues, "Quantum error correction well below the color code threshold". It is currently available as a preprint, and no journal or DOI has been confirmed, so it is not a peer-reviewed paper. As of October 20, no independent research group has been reported to have reproduced the result.
More qubits, fewer errors: a demonstration of the color code's effect
The research team compared the error rates of triangular color codes with code distances of 3, 5, and 7. The code distance expresses the strength of protection against errors. Quantum error correction is a technique that encodes information across multiple physical qubits so that logical information survives even if some qubits suffer errors. However, the operations used to detect errors are themselves noisy, so adding physical qubits does not always improve reliability. The system must meet a "below-threshold" condition, in which the benefit of stronger protection outweighs the errors introduced by the additional operations.
This study confirmed that the logical error rate falls as the code distance increases. The Tesseract decoding results listed in Table I of the paper are as follows.
| Code distance | Physical qubits per logical memory | Trials | Failures | Logical error rate per check round |
|---|---|---|---|---|
| 3 | 13 | 15,840 | 178 | 0.055% |
| 5 | 37 | 21,600 | 17 | 0.0038% |
| 7 | 73 | 29,000 | 3 | 0.0005% |
All of these are results from 20 rounds of error checking on the real Helios hardware, not estimates from simulation. Six logical memory experiments were run in parallel at distance 3 and two at distance 5, adjusted so that the total number of physical qubits used was roughly the same as in the distance-7 experiment. The error rates in the table are approximate values calculated with the method described above, and the distance-7 figure in particular retains statistical uncertainty because of the small number of observed failures.
The logical error rate dropped from 0.055% at distance 3 to 0.0038% at distance 5 and 0.0005% at distance 7. The team fitted the three measurements to a relationship describing exponential decay and evaluated the error suppression factor for each increase of 2 in code distance at 12. The uncertainty on this factor is +3 and −2, which does not mean the ratios between adjacent measurements were each exactly 12. What matters is that, within the range of code distances tested, the logical error rate was measured to fall as the protection of quantum information was strengthened.
Helios also incorporated measures against "leakage errors," in which a qubit's state escapes from the two states used for computation. In addition to a method that uses light to return ions to the computational states, the team combined a method that swaps the roles of data and measurement qubits within the circuit and removes leakage errors through measurement and reset. Quantum Computing Report's analysis likewise cites addressing leakage errors on both the hardware and circuit sides as a characteristic of this implementation.
The improvement in error rate therefore cannot be attributed simply to adding more physical qubits. It was achieved under experimental conditions that combined leakage countermeasures, syndrome extraction circuits, decoders, and more, and it cannot be concluded that any single technology alone brought the improvement.
Logical gate accuracy improves, but real-time decoding remains a challenge
The team evaluated not only how well the logical memory retains information but also the reliability of the quantum operations themselves, in a separate experiment. In tests using two-qubit Clifford operations, the logical error rate per operation was 0.07% at code distance 3 and 0.007% at distance 5. The distance-5 estimate is 7×10⁻⁵, with an error of 3×10⁻⁵ in both directions. This, however, is a measurement of errors that occur during logical operations, not of a memory that holds quantum information. The 0.0005% error rate from the memory experiment therefore cannot be applied directly to the accuracy of quantum operations.
In the logical operation tests, sequence lengths were set to 69, 139, and 209, and 2,220, 1,110, and 666 circuits were sampled, respectively. These were run unencoded and protected at code distances 3 and 5, and the logical operation error rates were compared. Two-qubit logical operations at distance 7 were not evaluated because Helios does not have enough physical qubits available. This is another instance of the constraint that stronger protection of logical qubits requires more physical qubits.
The team also examined the processing speed of the classical computation that estimates errors from qubit measurement results. They compared a convolutional neural network (CNN) with a Transformer-type decoder (RTT) that processes information from each check round. Both were trained on simulated data from a Helios emulator with noise strength set to 4 times the actual level, and no additional training was done when evaluating on real-hardware data.
The average decoding time at code distance 7, measured on an NVIDIA H100 80GB GPU, was 0.52 milliseconds per round for the CNN and 0.35 milliseconds for the RTT. Although the RTT processed faster, all of this decoding was done offline. The computation time was measured by feeding recorded hardware data in round by round, and it does not include communication with the quantum device or the delay of feedback control based on decoding results. It has therefore not been demonstrated that errors can be corrected at this speed while an actual quantum computation is running.
In addition, when the same distance-7 hardware data was analyzed, the CNN judged 3 trials to be failures, whereas the RTT judged 5. For a decoder to be built into practical quantum computation, the accuracy of its error estimates matters as well as its speed. A future challenge is whether the low logical error rate can be maintained in a real-time environment that includes communication with the quantum device and feedback control.
An estimate of 30,000 qubits to protect 100 logical qubits, and its assumptions
Based on the error suppression factor obtained here, Quantinuum has also estimated the number of physical qubits needed to reach a target logical error rate of 10⁻¹². According to the company, protecting 100 logical qubits would require about 30,000 physical qubits in the QCCD architecture, which performs quantum operations while moving ions around. By comparison, a superconducting color code implementation is estimated to need about 900,000, a difference of roughly 30 times. The corresponding code distances are given as 21 for QCCD and 109 for superconducting.
However, this comparison rests on an extrapolation that assumes the error suppression trend confirmed in this experiment continues at larger code distances. The company has not actually protected 100 logical qubits using 30,000 physical qubits, nor has it demonstrated a 30-fold efficiency gap over superconducting quantum computers in general. The number of physical qubits required alone also cannot be used to compare total system cost, including operating speed, control equipment, and cooling facilities.
Realizing general-purpose fault-tolerant quantum computation requires non-Clifford operations such as the T gate, in addition to the Clifford operations evaluated here. That in turn requires a mechanism for preparing special quantum states called "magic states" and using them in computation. Quantinuum itself explains that it has not yet demonstrated below-threshold scaling, in which the error rate falls as code distance grows, for magic state preparation.
Quantinuum aims to complete "Apollo," a general-purpose fault-tolerant quantum computer, in 2029. The result is significant in showing on real hardware that adding physical qubits to strengthen protection can sharply reduce the error rate of logical qubits. On the road to practical quantum computing, however, scaling up to many logical qubits, preparing magic states, and integrating control systems, including real-time decoding, remain to be done. Whether low error rates can be maintained through long, complex quantum computations that combine these technologies will be a key measure of future progress.