Research

Moth’s Tessa encodes 1,568 values on a 120-qubit chip

Moth’s Tessa recovered 1,568 real values from all 120 qubits of IBM’s Phoenix Nighthawk r2, but pair-mode value correlation fell to 0.69 and later use remains unproven.

Moth’s Tessa Encodes and Recovers 1,568 Values on a 120-Qubit Chip
Moth’s Tessa Encodes and Recovers 1,568 Values on a 120-Qubit Chipgeralt / pixabay
By James Calloway6 min read

By James Calloway

Moth announced on November 16, 2026, that its Tessa method encoded and recovered 1,568 real values across all 120 qubits of IBM’s Phoenix Nighthawk r2 processor. Daniel Bultrini’s paper says the method addresses a persistent obstacle for quantum machine learning (QML, learning methods that use quantum circuits): loading classical data (ordinary numerical information) into fragile qubits on sparsely connected hardware.

The result concerns data placement and recovery, not a completed learning task. That distinction matters because the paper’s record rests on a specific definition of what counts as recovered data.

What does the record actually include?

Bultrini defines the record around experiments that loaded freely chosen numeric lists onto physical quantum processors and then retrieved the individual values. Runs that produced only a derived quantity fall outside that definition.

The paper reports that QCrank previously recovered 128 values from 12 trapped-ion qubits. A separate Fez experiment recovered 1,024 pixels from 13 qubits, but the paper says that result is not directly comparable because it used a trained tensor-network approximation (a compact learned representation), rather than an arbitrary list of real values.

The paper therefore identifies Tessa’s Phoenix run as the largest reported result under that definition. The claim does not show that the method has solved quantum data loading across all workloads.

Takeaway: The record is about recovering chosen numbers from physical chips, not completing a quantum-machine-learning task.

How does Tessa use the chip’s wiring?

Tessa, short for Tessellated QPIXL, alternates data qubits and address qubits in a checkerboard over the processor’s physical connections. Each data qubit is controlled only by neighboring address qubits, and nearby data qubits can share those controls.

Every additional address neighbor doubles a data qubit’s storage. An interior data qubit with four neighbors therefore holds 16 values, according to the paper. The construction works on any bipartite coupling graph, where every hardware coupler (physical connection) joins a data qubit to an address qubit.

Using the device’s native layout avoids SWAP gates (operations that exchange qubit states), routing (moving operations between nonadjacent qubits) and ancilla qubits (extra helper qubits). Moth says Tessa loads arbitrary classical data across an entire chip “without SWAP gates or helper qubits.”

The paper proves that shared controls do not prevent independent recovery of each stored value. All values stored in one basis can then be decoded from one round of computational-basis measurements (a readout of the qubits’ 0 and 1 states) across the chip.

Tessa can also place a second number in each slot’s phase, an additional state property. In the Phoenix pair-mode experiment, that second field carried image hue while the first carried brightness.

Takeaway: Tessa fills the chip by using its wiring, without extra helper qubits or state-swapping operations.

What does pair mode cost?

Pair mode increases capacity, but the paper shows two penalties. It requires additional measurement bases and doubles circuit depth (the number of sequential gate layers). It also divides measurement shots (repeated observations) among local address configurations.

The paper says the number of shots needed for equivalent accuracy grows exponentially with local coupling degree, which measures how many address neighbors a data qubit has. The experiments kept that degree small, limiting the trade-off encountered there.

The 1,568-value Phoenix run placed brightness and hue values in 784 used slots. Spare slots held zeros so that every one of the processor’s 120 qubits participated.

Takeaway: Double encoding fits more values but costs more circuit work and measurements.

What did the physical hardware recover?

Single-basis tests used 8,192 shots per basis on the Phoenix and Fez processors. They recovered a 28-by-28 QCAT image from 784 of Phoenix’s 792 slots and a 21-by-21 image from 441 of Fez’s 448 slots.

The paper reports raw hardware value correlations, a measure of agreement between stored and recovered values, of 0.96 on Phoenix and 0.95 on Fez. Mean absolute errors, or average reconstruction deviations, were 0.130 and 0.131, respectively. No error mitigation was used.

In those tests, dynamic-range recovery reduced the mean absolute errors to 0.075 on Phoenix and 0.078 on Fez. The method uses summary statistics of the input rather than the input itself.

The denser Phoenix pair run was less accurate. It produced a value correlation of 0.69, a value mean absolute error of 0.152 and a hue mean absolute error of 0.092.

That run used 1,456 two-qubit CZ (controlled-Z) gates, compared with 728 in the single-basis Phoenix test. Full-chip retrieval therefore worked on two coupling graphs, but the phase-encoded value came with greater circuit demands and lower reconstruction quality.

Takeaway: The chips recovered images with errors, and the denser pair-mode run did noticeably worse.

Did the tests extend beyond images?

The experiments also processed audio, MIDI music data, DNA and protein amino-acid sequences, including p53. The paper says the biological sequences were selected for their lengths or names, not for their biology.

Those cases therefore tested data handling, not biological analysis. At 32,768 shots per measurement, the paper reports that Phoenix recovered the 784-nucleotide MT-CO3 DNA sequence with 92.1% of nucleotides exactly correct without post-processing. The figure reached 92.3% with the best tested dynamic-range recovery.

That best result came with an important caveat: the recovery method was selected using the true sequence. Recovery remained worse for densely packed protein sequences, for which the paper reports no numerical accuracy. The paper cautions that one incorrect nucleotide or amino acid can have serious biological consequences.

The figures do not show that either sequence was scientifically interpreted or that a protein was functionally validated. They measure how closely a quantum processor reconstructed chosen strings, with the known-answer selection limiting what the 92.3% result establishes.

Takeaway: The DNA and protein tests measure copying fidelity, not biological meaning.

Can outsiders inspect the work?

Moth says Tessa is available for image-processing experiments on its Atlas platform. The repository’s documentation provides the library, raw hardware counts, job records, simulation outputs and notebooks (executable analysis documents) used to regenerate the paper’s figures and tables. The materials can be run without an IBM account, according to Moth.

The run log dates the principal Phoenix and Fez QCAT experiments to September 26, 2026. Later genetic and pixel-art runs appear in October 2026, placing the hardware work before the November 16 announcement.

Those artifacts make the reported counts and analysis inspectable. Moth’s announcement does not report an independent replication, so the public files support verification of the published work without documenting outside validation.

Takeaway: Moth released files for inspection, but the announcement contains no independent replication.

What remains unproven?

Tessa’s immediate contribution is a hardware-aware way to load and verify classical data across an entire processor. The paper leaves open “whether subsequent computations can exploit that encoded state without paying for full reconstruction.”

That question separates a successful storage demonstration from a useful quantum workflow. If later circuits must reconstruct every value first, the encoding may add work rather than remove it. The paper also leaves shot growth and the errors of denser pair encoding unresolved.

Moth has demonstrated loading and recovery at the scale that leads the paper’s narrow record category. It has not yet shown that a later computation can use the state directly while avoiding full reconstruction and controlling both measurement costs and errors.

Takeaway: The next test is whether later circuits can use the encoded data without reconstructing everything first while keeping measurement costs and errors under control.

Source: unite.ai

Share this article:

More from James Calloway

James Calloway

Show full bio

News editor covering people, launches and day-to-day practice in artificial intelligence and machine learning at AI In Context.

722 articles

Related articles

  1. IBM's Nighthawk r3 hits 520 qubits and 19x speedup
  2. MIT Plugs GPT-5.6 Sol Into Quantum Lab, and It Calibrates Qubits Overnight
  3. Reimei Upgraded to 56-Qubit H2 as Quantinuum Widens RIKEN Deal
  4. Intel's $2.3 Billion Classiq Deal Signals Quantum Computing's New Phase
  5. A Quantum Physicist Is Using OpenAI's o1 to Tackle Physics' Biggest Questions

« Previous articleNext article »