On November 10, 2026, RIKEN announced that it had upgraded Reimei, the trapped-ion quantum computer it operates at its Wako campus, from Quantinuum's H1 to the H2. The system has 56 fully connected qubits, meaning any two qubits can be operated on directly. Alongside the joint research between the two organizations and the development of future quantum computers, a separate memorandum that includes SoftBank has also been announced.

What will the upgraded machine compute? At the center of the research is a division of labor: data obtained from the quantum hardware is passed to AI and supercomputers to predict the properties of materials and molecules. However, the research results presented so far are preliminary, and they do not show that a commercial service is ready.

AD

Dividing the Work Between Quantum Hardware and Supercomputers

Reimei is operated under JHPC-quantum, which is supported by NEDO. The figure of 56 qubits does not refer to logical qubits with error correction applied. "Fully connected" describes a property of the device: a two-qubit gate can be executed directly on any pair of qubits.

According to a technical explanation that Quantinuum published on November 10, 2026, Reimei will be used as an accelerator that forms part of a classical computing environment. The design, managed by the company's Nexus, has classical computing call the needed quantum computations and receive the results. Not every step is moved onto Reimei; work is divided with Fugaku and ROQUO, which is equipped with GPUs, depending on the target of the calculation.

Reimei handles quantum data and candidate electron configurations, ROQUO handles classical simulation and AI, and Fugaku handles the molecular environment and chemical calculations. The accuracy improvements are preliminary and are not measurements showing commercialization.

Computer Role in the materials model Role in zinc porphyrin
Reimei Acquires quantum data in regions that are hard for classical computing Searches for candidates of important electron configurations
ROQUO Classical simulation and AI training and inference Optimization and configuration interaction calculations, AI-based circuit design and spectrum reconstruction
Fugaku No individual step described Surrounding molecular environment and overall chemical calculations

The table organizes the steps assigned in Projects 1 and 2 of the explanation by study, and it does not rank performance. The technical explanation does not describe Fugaku's individual role in the materials model. Where quantum computation is built in and what it returns to classical computing are the keys to reading the two studies.

From a 6×6 Spin Model to Data for AI

In the materials research, data from Reimei's actual hardware has reportedly already been collected for a 6×6 two-dimensional spin model. This is computational data for a model, and it must be distinguished from measurements of real material samples.

In the study, regions that classical computing can handle are assigned to simulation on ROQUO using tensor networks, and difficult regions are assigned to Reimei. Quantinuum cites the growing burden on classical computing in regions with strong correlations and entanglement as the reason for this division. The design combines the two types of data obtained this way to train a generative AI.

Data from real hardware contains noise. In this study, a Transformer was used for "zero-noise extrapolation," which estimates the values that would be obtained without noise. In addition, active learning was used to decide under which conditions to preferentially collect additional quantum data.

After training, the AI is run on ROQUO to predict physical properties under conditions that have not yet been calculated. What these predictions fill in is the "phase diagram," which shows a material's properties according to conditions such as temperature and magnetic field. The step of collecting data from real hardware is separate from the step of obtaining values through simulation and AI.

RIKEN and Quantinuum reported that the initial results suggest the approach may predict more accurately than an AI trained on classical simulation alone, and more precisely than a quantum computer alone. However, actual figures and errors before and after the improvement have not been released, and the number of trials, the amount of training data and the evaluation metrics are not shown either.

There is not enough material to judge the size of the improvement numerically. This is a preliminary report comparing the predictions of computational models, and it is not a result that identifies the added quantum data alone as the cause of the improvement.

AD

How to Calculate the Light Absorption of Zinc Porphyrin

The other study is a biomolecule calculation by researchers from RIKEN, NVIDIA and Quantinuum. The initial target is zinc porphyrin, a molecule that absorbs light. The researchers investigate its light response, which depends on the surrounding environment, from its excited states, the high-energy electronic states involved in light absorption.

Here, the concept divides the work between calculations that include the molecule and its surroundings and calculations for regions where relationships between electrons are strong. Fugaku handles the surrounding molecular environment and the overall chemical calculation, while Reimei searches for candidates of important electron configurations. The candidates are passed to classical computing, which reconstructs energies and absorption spectra.

ROQUO handles optimization and configuration interaction calculations, along with AI-based quantum circuit design and spectrum reconstruction. It connects the step of searching for candidates on quantum hardware with the step of using those candidates to obtain energies and other quantities. This differs from an explanation in which the whole molecule is solved by a quantum computer alone.

The announcement also touches on the possibility that such calculations could be useful for solar cells and similar applications. However, what has been shown is a concept and initial efforts in computational research, not a result in which a solar cell was built and its improved performance measured.

At present, the two studies are reported through a press release and a corporate technical blog. No authors, journal names or DOIs for a corresponding peer-reviewed paper or preprint can be confirmed, and the official materials consulted contain no report of independent replication. Quantinuum's English-language announcement also adds a caveat: the results were obtained under preliminary, experimental conditions and may not be reproduced in larger-scale calculations, calculations with logical qubits, or commercial use.

Two Memoranda and the Conditions for Practical Use

In parallel with the upgrade, RIKEN and Quantinuum signed a memorandum on AI for Science and on joint development of a "Made with Japan" fault-tolerant quantum computer. The plan draws on the expertise of universities and research institutions in Japan, includes component and materials suppliers, and combines Quantinuum's technology with RIKEN's operating experience.

A fault-tolerant quantum computer refers to a machine that aims at large-scale computation by detecting and correcting errors during computation through quantum error correction. Completing the upgrade of Reimei to the H2 and completing such a machine are at different stages. Mitsuhisa Sato, director of RIKEN's division, said of the expected development of the research, "Progress in quantum-HPC hybrid computing is expected."

The three-party memorandum that includes SoftBank is a separate agreement. It will consider the possibility of commercial services, taking into account connections to AI data centers and the state of infrastructure development. It is not an announcement that commits to a launch date, pricing, form of service or completion date.

The technical explanation says regular operation of the Test User Program will begin after the two initial studies are finished. This transition, too, needs to be considered separately from the start of a commercial service for the general public.

The next materials for judging practical value are a comparison of prediction accuracy and required computing resources under the same conditions, and reproducibility in larger-scale calculations. If that verification is published, it will become possible to see concretely which calculations benefit from a division of labor that builds quantum hardware into AI and supercomputers.