Quantinuum said on September 8, 2026, that its researchers had tested Helix, the company’s error-correction architecture, on Helios, its trapped-ion processor with 98 physical qubits. In a preprint posted on September 2, 14 researchers at the company’s sites in Broomfield, Colorado, and London reported results on protected memory, logical Clifford gates, and entanglement between two different error-correcting codes.
In each of the three experiments, the authors wrote, the team beat its unencoded physical baseline with the encoded operation, without postselection. Postselection is the practice of discarding runs in which an error was detected.
Quantinuum’s team showed that Helix can store information and run Clifford gates, but the team decoded errors only after the runs and ran none of the non-Clifford gates that universal computation requires. The preprint has not been peer reviewed.
The code
Quantinuum designed Helix around what the authors call the [[20,2,6]] C4-Helix code. In that notation, 20 physical data qubits encode two logical qubits with a code distance of 6.
A logical qubit is a unit of information spread across many physical qubits so that errors can be detected and corrected. Code distance is the smallest number of physical errors that can change the stored information without being detected.
Quantinuum’s team built the code by concatenating a [[10,2,3]] code with a [[4,2,2]] code, nesting one code inside the other. The authors described it as a concatenated symplectic double code.
In a paper published in the journal Quantum on August 19, Noah Berthusen and Elijah Durso-Sabina, both affiliated with Quantinuum, wrote that codes of that type support many logical gates using only single-qubit gates and the relabeling of qubits.
Syndrome extraction is the repeated set of parity checks, performed by helper qubits called ancillas, that reveals errors without reading the stored information. The team used 32 physical ancilla qubits alongside the 20 physical data qubits to run those checks in parallel.
Three experiments
In the memory experiment, the team ran 20 rounds of syndrome extraction and measured 4.6 logical errors in 100,000 per logical qubit per round. In the preprint, the authors gave the 95% uncertainty as +6.2 and −2.6 in 100,000, a range of about 2.0 to 10.8.
The team measured only in the X basis, one of two standard measurement settings. The authors described that choice as a conservative proxy, given that dephasing is the dominant noise on Helios.
The error rate was about 4.5 times lower than the rate the team measured with the smaller [[10,2,3]] code alone. Quantinuum wrote in a blog post that, based on a study of existing literature, it believes the memory figure is a world record.
Clifford gates are a restricted set of logic operations that a classical computer can simulate efficiently. In randomized benchmarking, researchers estimate the average error of a gate set by running long sequences of randomly chosen gates.
With that method, the team benchmarked the complete set of Clifford operations on the code’s two logical qubits. The team measured 2.8 errors in 10,000 per two-qubit logical Clifford, about 4.3 times lower than Helios’s physical two-qubit Clifford error of about 1.2 in 1,000.
In adaptive syndrome extraction, the control system uses earlier check results to select which later checks to run. Using the method, the team reduced the number of physical two-qubit gates for one logical phase gate by about 33% and the average run time per shot by about 23%, the authors reported in the preprint.
In the third experiment, the team entangled two Helix logical qubits with one logical qubit stored in a distance-5 surface code, a different type of code. In a GHZ state, several qubits are entangled so that their measurement outcomes are perfectly correlated.
Fidelity is a measure of how closely a prepared state matches the ideal one, where 100% is perfect. In the preprint, the authors reported a fidelity of at least 99.925% for the logical GHZ state, against 99.537% for the same state on physical qubits. The authors wrote that they bounded the error of the entire sequence, from state preparation to readout, and did not isolate the link between the codes.
Magic states and the second code
A fault-tolerant computer uses error correction to keep errors in check as it computes. Magic states are specially prepared quantum states that such a computer consumes to perform non-Clifford gates. Quantinuum wrote in its blog post that the Helix code is “sub-optimal for preparing magic states.”
Quantinuum assigned that job to a surface code in its design, according to the post. The authors proposed preparing magic states in a surface or color code and passing them to the Helix code through a link like the one the team tested, which they call a chain map.
What the team did not show
A decoder is the classical software that works out, from the parity-check results, which errors occurred. The team decoded its data with a decoder called Frontier after the experiments had finished. The authors left for future work a streaming version that runs alongside the quantum computer.
The authors wrote that “the present experiment does not implement a universal Clifford+T computation.” In Clifford+T, the T gate is the non-Clifford gate.
From circuit-level simulations, the authors projected logical error rates between 1 in 1 million and 1 in 100 million if physical error rates improve to around 1 in 10,000. The team did not measure those rates on Helios.
Next steps
Quantinuum described Helix as a foundation for Apollo, the universal, fully fault-tolerant computer it has said it will deliver by 2029. The company wrote that it expects the same code to perform several orders of magnitude better on Apollo with higher physical fidelity.
Update, November 12, 2026: On October 19, Quantinuum reported color-code results on Helios, with a memory error of about 5 in 1 million per cycle at distance 7 and a two-qubit logical Clifford error of about 7 in 100,000. Both are lower than the Helix figures, though the two experiments used different codes and resources.
