Quantinuum Reports 99.9995% Logical Memory Fidelity on Helios

Quantinuum has reported a quantum error-correction result that brings a central requirement of useful quantum computing into sharper focus: keeping quantum information intact long enough to do something valuable with it. On its Helios trapped-ion computer, the company measured logical memory fidelity of 99.9995% per error-checking cycle—equivalent to about five logical failures per million cycles.

Announced on October 19, the result represents a tenfold reduction from the company’s previous reported record. Quantinuum says it strengthens the case for Apollo, its planned universal fault-tolerant system targeted for 2029. The experiment is a significant memory benchmark, however, rather than a demonstration of a complete useful application running at that accuracy.

The accompanying research also highlights a growing connection between AI and quantum computing: neural networks can help interpret error measurements quickly enough to support future real-time correction. Understanding what was measured—and how those decoders were tested—is essential to assessing the advance.

What the 99.9995% Result Measures

Physical qubits are vulnerable to disturbances that corrupt their quantum states. Error correction spreads information across multiple physical qubits to create a logical qubit, while additional measurements reveal signs of errors without directly reading out the protected information.

The key test is whether adding more protection actually reduces the logical error rate. Below an error-correction threshold, larger codes should perform better. Above it, the extra operations needed to maintain the code can introduce too much noise.

In Quantum error correction well below the color code threshold, Quantinuum researchers tested color codes at distances three, five, and seven. Code distance describes the strength of the protection. The measured logical memory error rate fell from approximately 5.5 in 10,000 per cycle at distance three to 3.8 in 100,000 at distance five, and about five in a million at distance seven.

The largest memory experiment used 73 physical qubits, including measurement ancillas, to protect one logical qubit. Its estimate came from three failures across 29,000 experimental shots, each containing 20 correction rounds. The small number of observed failures means statistical uncertainty matters: the paper reports a central error estimate of five per million with a one-standard-deviation interval extending approximately from three to nine per million.

That is strong evidence of increasingly effective protection. It does not mean that a million arbitrary quantum gates, or a million steps of a commercial algorithm, have been executed without error.

Why Helios Can Test Different Error-Correction Codes

Quantinuum’s technical account of the experiment emphasizes improvements made on the same Helios hardware used for its earlier result. The platform uses trapped ions, with connectivity that allows interactions beyond a fixed nearest-neighbor layout. That flexibility lets researchers change error-correction strategies without redesigning the processor around a single code.

The new work also addresses leakage, in which an ion leaves the states used to represent a qubit. Leakage can create correlated errors that are particularly damaging to correction schemes. The researchers combine optical repumping, which returns leaked ions to the computational states, with circuit techniques designed to limit how errors spread.

Across the tested code distances, the logical memory error rate improved by a factor of roughly 12 each time the distance increased by two. A separate two-qubit Clifford-gate experiment demonstrated suppression by a factor of about nine between distances three and five. These are related but distinct benchmarks: the headline memory fidelity should not be applied to the gate result.

AI Helps Decode the Error Signals

Quantum error correction depends on a classical decoder that interprets measurement patterns and determines the likely errors. Better physical hardware alone is insufficient if that interpretation cannot keep pace with the processor.

The study compares several approaches, 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 then applied to experimental data without hardware-data fine-tuning.

This is a specific role for AI: identifying error patterns in quantum measurements. The models are not replacing the quantum processor or demonstrating a general-purpose AI application. Their value is in balancing decoding accuracy with the speed needed to sustain repeated correction cycles.

The researchers report neural inference times in the hundreds of microseconds, compatible with real-time operation on this class of hardware. Crucially, all decoding in the study was performed offline. Streaming tests supplied the measurements round by round to assess timing; they were not a demonstration of the neural decoders controlling a live feedback loop throughout the experiment.

That distinction leaves an engineering milestone ahead while preserving the importance of the result: models trained in simulation successfully interpreted errors from a physical quantum machine.

What Remains Between a Memory Record and Useful Computing

Quantinuum describes the result as a world record and says its accuracy is now within the target range for Apollo. Those claims should be read alongside the broader requirements of fault-tolerant computing. A practical system must protect many logical qubits, execute a sufficiently broad set of gates, sustain correction during computation, and provide enough throughput to complete valuable workloads.

Even within this study, resource limits remain visible. The researchers did not run the distance-seven two-qubit gate experiment because it exceeded Helios’s available qubit count. The company’s blog also identifies scalable, below-threshold magic-state preparation—needed for an important route to universal fault-tolerant computation—as work still to be demonstrated.

Quantinuum’s roadmap includes its Sol system in 2027 and Apollo in 2029. Those remain development targets. The present evidence is narrower but meaningful: on existing hardware, changing the code, controlling leakage, and improving decoding substantially reduced logical memory errors.

For the quantum industry, the result shifts attention toward a more demanding question than whether a processor can demonstrate error correction at all: whether that protection can scale across the full machinery of useful computation. For AI researchers, it offers a concrete example of machine learning contributing to that effort at the level of the control infrastructure.