Moth sets quantum data encoding record on 120-qubit IBM chip

Akhil Ponnada
Akhil Ponnada is the founder and editor of FrontWord. He covers artificial intelligence, chips and infrastructure, consumer technology, cybersecurity and the business of tech.
6 Min Read
Image: TechRadar

Moth Quantum announced a new quantum data encoding method called Tessa on 16 November 2026, reporting that it loaded and retrieved 1,568 numerical values across an IBM 120-qubit processor. The test on IBM’s Phoenix chip retrieved arbitrary data points from physical qubits, addressing a hardware bottleneck in quantum machine learning.

According to Unite.AI, the experiment sets a record for loading arbitrary lists of numbers onto a physical quantum computer and reading them back out. Previous tests under the same definition reached 128 values on a 12-qubit trapped-ion system.

Getting classical data into fragile quantum systems has long slowed progress in quantum machine learning. In a research paper describing the project, Moth researcher Daniel Bultrini defined the benchmark strictly around arbitrary numeric arrays. While earlier projects loaded 1,024 pixels on a 13-qubit machine using trained tensor-network approximations, Tessa loaded arbitrary values directly.

How the Tessa method stores data

Tessa, short for Tessellated QPIXL, arranges data and address qubits in a checkerboard pattern mapped directly to the processor’s physical connections. Each data qubit links only to neighbouring address qubits, and adjacent data qubits can share those controls.

Every additional address connection doubles the capacity of a data qubit. An interior qubit with four neighbours can hold 16 separate values. Because the method fits the natural bipartite layout of the chip, it avoids routing signals across the processor and eliminates the need for helper qubits.

Bultrini designed the circuits to keep hardware demands low. The tested circuits reached a depth of 19 two-qubit operations, as reported by TechRadar.

“Tessa makes the most of the chip’s capabilities, so the whole processor can be put to work at once,” Bultrini said.

However, the technique faces scaling limits. Because measurement shots are split among local address configurations, the number of shots required for accurate recovery grows exponentially as the number of local connections increases.

Accuracy limits on physical hardware

Reading data out of physical qubits introduces noise. In single-basis tests, researchers used 8,192 measurement attempts to reconstruct a 28-by-28 pixel image on Phoenix and a 21-by-21 image on IBM’s Fez processor.

Those single-basis runs achieved correlation scores of 0.96 on Phoenix and 0.95 on Fez. Mean absolute numerical errors stood at 0.130 and 0.131. Applying dynamic-range recovery based on summary statistics from the original image lowered the errors to 0.075 and 0.078.

To reach the figure of 1,568 values, Moth used a pair mode that placed two numbers into each slot by encoding one in the qubit’s phase. Phoenix ran 784 active slots containing both brightness and hue data, using 1,456 controlled-Z gates.

Packing two numbers into each slot reduced fidelity. Value correlation dropped to 0.69, while the mean absolute error rose to 0.152 for brightness and 0.092 for hue. The brightness error on Phoenix was 2.6 times higher than predicted by Moth’s noise model, pointing to unmodelled physical hardware noise.

None of the hardware runs used quantum error mitigation or error-correction techniques. The researchers noted that applying error mitigation could significantly cut reconstruction errors in future trials on physical processors.

Tests across audio, music and genetics

The team also tested non-visual data formats to evaluate how the system handles different information structures. A test using 347 flute notes recovered 18.7% of pitches correctly on raw hardware, rising to 49.3% after post-processing adjustments on the Fez chip.

Biological data tests demanded 32,768 measurement attempts per run. The system recovered 92.1% of nucleotides correctly on a 784-unit mitochondrial DNA sequence without adjustment, and 92.3% with dynamic-range recovery.

More complex protein structures struggled. The p53 protein sequence reached 71.5% accuracy and Sonic hedgehog reached 76.2%, but pikachurin managed only 25.8% and SUR1 scored 22.6%. The authors noted that biological systems tolerate little error, meaning these recovery rates remain impractical for medical work.

What happens next with the data

The physical hardware runs took place in September and October 2026, ahead of the formal announcement in November, as Moth announced. Harry Kumar, Moth’s founder and chief commercial officer, said high-impact quantum applications linked to artificial intelligence depend on solving data-loading limits.

The company released its code, job logs and raw counts in its repository documentation, enabling outside researchers to reproduce the paper’s charts. Moth has made Tessa available for image-processing experiments on its Atlas platform.

The next question for researchers is whether quantum machine learning algorithms can perform calculations directly on the encoded quantum state without paying for full reconstruction.

Follow:
Akhil Ponnada is the founder and editor of FrontWord. He covers artificial intelligence, chips and infrastructure, consumer technology, cybersecurity and the business of tech.