A quantum computer can only act on information once that information has been carried into it, and that carrying-in has long been a quiet bottleneck. Moth Quantum now reports that its technique, Tessa, loaded 1,568 numbers across all 120 qubits of IBM's Phoenix processor and brought them back out, a claimed record for this kind of test. The recovered images show the cost of the feat, with the most complex picture coming back only loosely faithful to its source. Whether this becomes a real doorway for the machines of tomorrow depends on tests the field has not yet run.
Moth Quantum's Tessa Stores 1,568 Values on IBM's 120-Qubit Processor, But Image Recovery Remains Imperfect
The hardware misbehaved beyond what the noise model predicted
Start with the headline number. Why does 1,568 values matter?
Because loading data is the unglamorous first step. The earlier comparable result was 128 values on 12 trapped-ion qubits, so this uses a much bigger chip and far more data.
Worth noting that's a record by the study's own definition, and it's Moth's claim. I don't see independent verification in the reporting.
What's actually new in the method?
Tessa uses neighboring qubits to steer where information goes, so it relies on connections the chip already has. No extra helper qubits, no moving data between distant qubits.
And the images came back fine?
Simpler ones did, around 0.95 to 0.96 correlation. The largest color image only hit 0.69.
And the improved error figures, 0.075 and 0.078, used two statistics from the original image. That's partly cheating with known information, so I wouldn't read them as pure recovery.
Anything the company didn't predict?
Phoenix's brightness error was 2.6 times what their noise model expected. The hardware misbehaved beyond the model.
The biology results are the widest spread: 92.3% for MT-CO3, 22.6% for SUR1. Nothing here supports medical use, and the study says so.
So what would change the picture?
Error mitigation, which they didn't use. The study says it could reduce errors significantly.
Could, in future tests. That hasn't been shown yet.
The Pulse
- Moth Quantum says its Tessa technique pushed 1,568 values through IBM's 120-qubit Phoenix processor and pulled them back out, a claimed record.
- The loading step that every quantum machine needs has been a stubborn obstacle, especially for quantum machine learning.
- The largest color image returned with only 0.69 correlation to its source, while simpler images held near 0.95, exposing where the strain shows.
- Audio and DNA tests swung widely, with flute pitches recovered at 18.7% before adjustment and some protein sequences falling as low as 22.6%.
- Results lean on adjustments drawn from the original images and no error mitigation was applied, so the next test on error-corrected hardware is the real verdict.
A quantum computer can only act on information once that information has been carried into it, and that carrying-in has long been a quiet bottleneck. Moth Quantum now reports that its technique, Tessa, loaded 1,568 numbers across all 120 qubits of IBM's Phoenix processor and brought them back out, a claimed record for this kind of test. The recovered images show the cost of the feat, with the most complex picture coming back only loosely faithful to its source. Whether this becomes a real doorway for the machines of tomorrow depends on tests the field has not yet run.
Moth Quantum says it has set a record for a deceptively simple task: getting ordinary data into a quantum processor and getting it back out again. Its technique, Tessa, placed 1,568 numerical values across all 120 qubits of IBM's Phoenix chip, then rebuilt an image from the numbers it recovered. Daniel Bultrini, the Moth researcher who developed Tessa, explained that every quantum computer must first take in its data, and Tessa is designed so the whole processor works at once.
Earlier work of this kind recovered 128 values using 12 trapped-ion qubits, so the new figure is substantially larger by the study's own measure. Tessa exploits the chip's existing neighboring-qubit connections to control where information is stored, avoiding costly moves between distant qubits and the need for extra helper qubits.
The image results reveal the limits. Simpler pictures came back with correlations of 0.95 to 0.96, but the largest color image reached only 0.69. An adjustment improved accuracy, yet it relied on statistics taken from the original image, so part of the gain came from information already known. The study also cautions that the hardware added inaccuracies its noise model did not predict, and that biological results, though strong for some sequences, are not yet dependable enough for medical use.
Moth applied no error mitigation. Its researchers suggest such methods could significantly reduce errors in future hardware tests. Harry Kumar, Moth's Founder and CCO, argued that the most consequential quantum applications, including those tied to AI, will depend on breakthroughs like this. The open question is whether tests on error-corrected hardware can close the gap between the headline number and the quality of what comes back out.
Moth Quantum says it has set a record for a deceptively basic job in quantum computing: putting ordinary data into a quantum processor and getting it back out. Its technique, called Tessa, placed 1,568 numerical values across all 120 qubits of IBM's Phoenix processor, then used the recovered numbers to rebuild an image.
The company frames this as a bottleneck. Before a quantum machine can do anything useful with classical data, that data has to be loaded, and the step has been a particular obstacle for fields such as quantum machine learning. Daniel Bultrini, the Moth researcher who developed Tessa, put it this way: every quantum computer has to get data in first, and Tessa uses the chip's capabilities so the whole processor can work at once.
For scale, earlier work of this kind recovered 128 values using 12 trapped-ion qubits. By the study's own definition, the new result is substantially larger. Tessa uses neighboring qubits to control where information is stored, so the processor's existing connections do the work. That avoids operations that move information between distant qubits and avoids extra helper qubits. The tested circuits reached a depth of 19 two-qubit operations.
The image tests show where the limits lie. Two experiments stored 784 and 441 values, needing 728 and 428 controlled operations respectively, and each used 8,192 measurement attempts. The largest color image reached a correlation of 0.69 with its source, while simpler images reached 0.95 to 0.96. The average numerical error was 0.130 on Phoenix and 0.131 on IBM's Fez processor. After an adjustment, those fell to 0.075 and 0.078. But that adjustment used two statistics taken from the original image, so the improvement leaned partly on information already known. In a separate test combining brightness and color, Phoenix recorded errors of 0.152 and 0.092, and Fez 0.144 and 0.084. The Phoenix brightness error came in 2.6 times higher than the company's noise model predicted, a sign the real hardware added inaccuracies the model missed.
Moth also tried audio, musical notes, DNA and protein sequences, each represented inside the processor in different ways. A test of 347 flute notes recovered 18.7% of pitches correctly before adjustment and 49.3% after, on Fez.
The biological tests used 32,768 measurement attempts, four times the 8,192 used for images, yet accuracy still varied substantially between sequences for every measurement basis. On Phoenix, the p53 protein sequence was recovered 71.5% correctly, Sonic hedgehog 76.2%, and the mitochondrial protein MT-CO3 92.3%. Others fared far worse: SUR1 reached 22.6% and pikachurin 25.8%, suggesting longer or more complex sequences stay hard. The tests also covered a mitochondrial DNA sequence of 784 building blocks and a POLG2 messenger RNA sequence of 1,582.
The study cautions that a single wrong DNA or protein unit can matter biologically, so these percentages do not establish dependable medical uses. It lists other caveats: the simulations omit some interference between neighboring processor groups, the hardware runs used no error-correction techniques, and Tessa depends on a particular pattern of qubit connections, with measurement demands rising sharply as those connections multiply.
Moth did not apply error mitigation, which the study says could cut reconstruction errors significantly in future hardware tests. Harry Kumar, Moth's Founder and CCO, said the quantum applications with the greatest impact, including those working closely with AI, will depend on breakthroughs like Tessa. The next thing to watch is whether tests on error-corrected or mitigated hardware close the gap between the headline number and the quality of what comes back out.
Notable Quotes
Every quantum computer has to get data in before it can do anything useful with it, and that step has been a bottleneck for fields like quantum machine learning.— Daniel Bultrini, Moth Quantum researcher who developed Tessa
The quantum applications with the greatest impact, including those working closely with AI, will depend on breakthroughs such as Tessa.— Harry Kumar, Moth Quantum Founder and CCO (paraphrased)