Microsoft and Quantinuum have demonstrated 12 logical qubits built from 97 physical qubits, running fault-tolerant algorithms that outperform the raw hardware beneath them. The achievement, detailed in a preprint released earlier this year, represents the largest error-corrected quantum circuits executed to date and shows that useful computations can survive the noise that has long crippled quantum processors. For researchers and companies betting on quantum computing, this result narrows the gap between laboratory demonstrations and machines that can solve real problems.
What is verified so far
Two preprints anchor the technical claims. The first, titled “Demonstration of logical qubits and repeated error correction with better-than-physical error rates,” reports that logical qubits achieved error rates below their physical baselines, with improvements reaching roughly 800 times when post-selection filtering was applied. That work, available on arXiv, established entanglement between logical qubits and showed repeated rounds of error correction, meaning errors were caught and fixed in real time rather than accumulating until the computation collapsed.
The second preprint builds directly on that foundation. Titled “Fault-tolerant execution of error-corrected quantum algorithms,” it reports running two well-known algorithms, QAOA and HHL (applied to a Poisson problem), on Quantinuum’s H2 and Helios trapped-ion processors using the Steane error-correcting code. The circuits used 12 logical qubits assembled from 97 physical qubits, as detailed in a separate preprint, making them the largest fault-tolerant quantum circuits reported so far. Both algorithms completed with error-corrected outputs, not just memory benchmarks or idle-qubit tests.
The distinction matters. Earlier demonstrations proved that logical qubits could sit in memory longer than physical ones or that entangled pairs could be preserved. Running full algorithms through multiple rounds of gates and corrections is a harder test because every operation introduces fresh chances for failure. The fact that QAOA and HHL ran successfully on error-corrected hardware means the Steane code handled not just storage errors but also gate-level faults during active computation. In other words, the logical layer is no longer a passive shield; it is actively stabilizing a live calculation.
Quantinuum’s broader research trajectory, documented in related preprints hosted on arXiv’s platform, includes an accelerated roadmap targeting universal, fully fault-tolerant quantum computing by 2030. That timeline appears in a separate technical paper and has not been independently validated by third-party benchmarks or replication studies. It functions more as an internal goalpost than as an industry-wide consensus forecast.
What remains uncertain
Several gaps separate these preprint results from a production-ready fault-tolerant machine. The 800-times improvement figure relies on post-selection, a technique that discards runs where errors were detected rather than correcting them in place. Post-selection inflates success rates by keeping only the best outcomes, which means the effective throughput of the processor drops with each discarded run. A system that delivers highly accurate answers but has to throw away most attempts may still be valuable for research, yet it is not the same as a high-volume, always-on computing service.
Without post-selection, the error suppression would be smaller, though neither preprint provides a direct comparison at the same circuit depth and qubit count. That missing side-by-side view makes it difficult to estimate how quickly post-selection could be phased out as hardware and decoding algorithms improve. It also complicates comparisons with alternative error-correcting codes and architectures that emphasize inline correction over aggressive filtering.
Raw experimental data and hardware logs from the H2 and Helios runs have not been released beyond what the preprints summarize. Independent teams have not yet replicated the results on different hardware platforms, and the preprints have not completed formal peer review. Until those steps occur, the claims rest on a single collaboration’s measurements and analysis. Microsoft’s specific engineering contributions, such as software stack design, decoder optimization, or calibration protocols, are referenced through citations and acknowledgments rather than detailed in a standalone technical disclosure, leaving some implementation choices opaque.
The 2030 roadmap for universal fault tolerance is a projection, not a guaranteed schedule. It depends on scaling trapped-ion systems to far larger qubit counts while maintaining or improving the per-gate error rates demonstrated in the current experiments. Trapped-ion platforms are known for high-fidelity gates but face challenges in control complexity and system size as more ions are added and interconnected. The roadmap assumes these engineering hurdles can be managed without degrading the delicate error-corrected operations seen in small-scale experiments.
Whether the Steane code, which uses seven physical qubits per logical qubit at distance three, can scale efficiently to the thousands of logical qubits needed for commercially relevant problems like molecular simulation or cryptanalysis is an open engineering question. Higher code distances would improve protection but multiply the physical qubit overhead and the number of gates per logical operation. At some point, the extra complexity risks erasing the gains from error correction unless hardware fidelities and decoding algorithms improve in lockstep. Competing approaches, such as surface codes, may offer different trade-offs, but the current work does not resolve which code family will dominate at scale.
How to read the evidence
The strongest evidence comes from the two preprints themselves. Both are primary technical documents with named authors, described methods, and quantified results. They report what was measured on specific hardware, not what a model predicts or what a press release promises. Readers should weight these claims more heavily than roadmap announcements or corporate timelines, which describe aspirations rather than completed experiments. At the same time, preprints are early-stage communications, and their conclusions can shift under peer review or replication.
The 97-physical-qubit figure is concrete and verifiable: it describes the actual hardware footprint of the 12-logical-qubit circuits. The 800-times improvement is real but conditional on post-selection, so it represents a best-case scenario rather than a steady-state operating metric. Both numbers appear directly in the preprint abstracts and can be checked against the full papers. Interpreting them correctly means recognizing that “largest so far” and “better-than-physical error rates” are milestones on a continuum, not endpoints.
Context from Quantinuum’s earlier publications, including studies on error correction thresholds and trapped-ion gate fidelities, supports the plausibility of the new results but does not independently confirm them. These papers form a citation chain showing incremental progress, not a single leap. The arXiv repository, maintained as an open-access service, provides visibility into this progression but does not perform peer review or experimental validation, so external scrutiny remains essential.
For anyone tracking quantum computing as a technology investment, research direction, or competitive factor, the practical takeaway is specific. The Microsoft–Quantinuum experiments show that double-digit logical qubit counts, running nontrivial algorithms with demonstrable error suppression, are now within reach on a commercial trapped-ion platform. They do not show that large-scale, general-purpose quantum computers are imminent, nor that error correction has been “solved” in an economic sense. The field has crossed an important technical threshold, but the distance from these carefully curated demonstrations to broadly useful, fault-tolerant services is still measured in years, engineering breakthroughs, and many more qubits.
More from Morning Overview
*This article was researched with the help of AI, with human editors creating the final content.
