C12 and Thales Win the 2026 Quantum Effects Hardware Award for QuantumTrack, a Radar Multiple Hypothesis Tracker: What the April arXiv Paper Shows and the TRL 5 Label Does Not
Quentir Defense Monitor
Evidence-based insights for quantum defense and security. Published by Quentir Systems LLC · September 29, 2026.
On September 29, 2026, the Paris spin-qubit company C12 and Thales announced QuantumTrack, a hybrid quantum-classical method for real-time multi-target radar tracking, and Messe Stuttgart named it the winner of the 2026 Quantum Effects Award in the quantum computing hardware category. The release says the method matched the time-to-solution of the strongest classical solver on its benchmark, ran about 100 times faster than competing quantum annealers, has reached Technology Readiness Level 5, and is headed for a TRL 6 demonstration on C12's physical processor with defense systems as the target.
The checkable evidence behind those sentences is a short paper posted to arXiv on April 16, 2026 by Quentin Schaeverbeke and Viktor Radović of C12 Quantum Electronics and Jean-Marc Divanon and Bing Hong Teh of Thales Research and Technology in Singapore. The paper simulates a quantum annealer built from carbon nanotube spin qubits inside a microwave resonator, runs the annealer on two small synthetic tracking scenarios, and estimates that a full run could take about 50 milliseconds once the processor supports active reset and parallel readout. This post reads the paper against the release, places the result on the readiness ladder, and asks what a radar program office could do with it.
What the April 2026 paper by Schaeverbeke, Radović, Divanon and Teh actually ran: a two-target and a one-target scenario on the Callisto emulator
The problem is old and well posed. A surveillance radar produces a set of detections at every scan, some from real targets and some from clutter, and the tracker must decide which detection belongs to which track. Donald Reid formalized multiple hypothesis tracking in the late 1970s: keep several candidate assignments alive, score each by likelihood, and prune the improbable ones as later scans arrive. The count of hypotheses grows exponentially, the data association problem is NP-hard, and every fielded tracker prunes hard to hold its update rate. The paper's Figure 3 shows the effect on a toy case: with two objects moving at constant velocity and a clutter density of one in a hundred thousand per unit volume, the number of live hypotheses climbs by nearly three orders of magnitude between scan 4 and scan 10.
The step the partners hand to the quantum processor is the pruning itself, written as a Maximum Weighted Independent Set problem. Each hypothesis is a graph vertex weighted by its likelihood; two hypotheses that share a measurement are joined by an edge; the tracker wants the heaviest set of vertices with no edge between them. That problem maps directly onto an Ising Hamiltonian, and quantum annealing looks for its ground state. C12 describes the same construction on its own case-study page, where it notes that the emulator was extended with a time-dependent annealing model, an Ising formulation of the independent-set problem, spin-qubit noise processes and a graph decomposition that splits instances of several hundred vertices into pieces the emulator can hold.
Two benchmark scenarios appear in the paper. In the single-step case, the classical Two-Stage MaxSAT solver for maximum weighted clique runs the whole track and identifies scan 10 as the moment with the most hypotheses; the emulated annealer then solves that one instance, and Figure 4 overlays the classical and emulated tracks against the ground truth. In the sequential case, a single target is tracked with the emulated annealer solving the pruning step at every scan, at two clutter densities. At the higher density the tracker prunes too eagerly and the recovered trajectory fragments; at the lower density it recovers the complete track. That is the entire experimental record: two objects in one scenario, one object in the other, synthetic measurements, and an emulator that the paper says can simulate up to 20 qubits in gate mode and was extended with a quantum Monte Carlo routine for the annealing runs.
How the 50 millisecond figure is built: 50 to 150 microsecond anneals, a 1 microsecond cQED readout, and an active reset not demonstrated on the processor in the cited paper
Real-time tracking lives or dies on latency, so the paper's most useful contribution is its accounting of where the time goes. Three parts add up: state preparation, the anneal itself, and readout. The anneal is the part the emulator can tune. Too fast and the system crosses energy gaps diabatically, producing coherent errors; too slow and decoherence from Purcell loss, charge noise and nuclear spin noise degrades the answer. Balancing the two, the authors find an optimal anneal between 50 and 150 microseconds. Readout on a circuit quantum electrodynamics platform is estimated at 1 microsecond per measurement, and if every qubit can be read in parallel, the whole register costs the time of one.
State preparation is where the estimate leaves the measured world. On the C12 hardware, the paper reports a passive reset time of 5 to 10 milliseconds, and a run that repeats the anneal for the required number of shots then lands on the scale of seconds. The 50 millisecond figure assumes an active reset, in which each qubit is measured and flipped back to its ground state by a conditional gate, cited from published work on silicon spin qubits rather than demonstrated on C12's device in the cited paper. The abstract states the result correctly as something cQED spin processors could reach. The conclusion is equally plain: with passive reset the total is about 5 seconds, and with active reset and parallel readout it falls to about 50 milliseconds in the most favorable case. A reader who saw only the release's line that real-time tracking is within reach would not know that the difference between a second and a twentieth of a second rests on a reset protocol the partners have identified but not demonstrated on C12's device in the cited paper.
Quantum pillar: computing (electronic warfare processing). Use posture: dual-use. Technology readiness: TRL 3 of 9. The tracker ran as working software on synthetic radar detections inside an emulator carrying a hardware-calibrated noise model, with no run on a physical processor reported, which on the shared ladder is a rung 3 result whatever label the release attaches to it.
Who gains from a real-time hypothesis pruner: air-defense radars facing saturation attacks, drone traffic management, and why the capability map reads it as dual-use
Under the capability map this lane uses, the sub-branch is signal separation in the radar processing chain: telling target returns from clutter and keeping the right association as targets cross and maneuver. That is the same computational choke point whether the radar is a civil airport sensor managing drone traffic or a naval air-defense set facing a saturation raid, and Thales's radar center of excellence director Jean-Marc Divanon, a co-author of the paper, says as much in the release when he names dense drone environments and emerging threats in one breath. The posture is dual-use because the capability serves whoever owns the radar. For a defender it means more hypotheses can survive each scan, so a tracker is less likely to drop a real target when dozens of decoys arrive at once. For an adversary who fielded the same tracker on a targeting radar, the gain is symmetric.
What the sources do not establish matters as much. Nothing in the paper measures tracker quality against the pruning a classical tracker already performs in the time available; the comparison is between an emulated annealer and one exact classical solver on tiny instances, and the 100-times claim against other quantum annealers appears in the release alone. The interesting engineering claim is architectural rather than a speed record. C12's spin qubits sit in carbon nanotubes coupled to a shared microwave resonator, which gives long-range interactions among qubits, and the same electrostatic gates that switch a qubit between memory and operating modes let the processor act as a native annealer. Semiconductor fabrication compatibility and a small footprint are the assets Thales would care about if the processor were ever to travel with a radar.
That last condition is the practical one. A radar tracker runs at the update rate of the antenna, and the Monitor has examined how the timing reference itself becomes a limiting resource when two cold-atom clocks hold a radar network together without satellites. A quantum annealer bolted to a radar would need a cryostat, a control stack and a latency budget inside the scan interval, none of which the paper addresses because it addresses the physics of the anneal. The annealing approach itself is portable across domains; the same formulation ranked drug combinations when a Porto and DLR team ran simulated quantum annealing on network medicine, and the lesson there was the same: the optimization step is well matched to annealing, and the surrounding system decides whether the match is worth anything.
What stands between the Quantum Effects award and a program office: no hardware run, two targets, a benchmark absent from the paper, and a TRL 5 label the shared ladder does not support
The first gap is the hardware. Every result in the paper comes from Callisto, an emulator with a noise model calibrated to C12's device. The physical processor contributed one measured number, the passive reset time, and the anneal, readout and active reset figures are estimates or citations. On the readiness ladder the Register uses, an algorithm running as working software on synthetic data is rung 3; a bench assembly under laboratory test is rung 4; trials with realistic targets are rungs 5 and 6. The release calls the project TRL 5 and promises a TRL 6 demonstration on real hardware next. Both companies are entitled to their own scale, and the Quantum Computing Report summary of the award repeats the 50 millisecond target latency as the project's defining number. A buyer should read the label as a statement of the partners' plan and the paper as the evidence, and the evidence sits two rungs lower.
The second gap is scale. Two objects and one object, in straight lines at constant velocity, with synthetic clutter, is the smallest possible exercise of a tracker whose whole justification is dense airspace. The release speaks of thousands of detections per scan and of decomposing large instances into subproblems sized for the processor; the case-study page mentions several hundred vertices handled through decomposition. How solution quality degrades when a hypothesis graph is cut into smaller pieces and stitched back together is exactly the question a program office would ask, and the public record does not yet answer it. The same question has followed every small-instance quantum optimization result the Monitor has read, including the IonQ and Capella synthetic aperture radar change-detection run on 20 qubits earlier this month, which reported an F1 score a classical baseline could match.
The third gap is the benchmark itself. The release says QuantumTrack matched the strongest classical solver's time-to-solution and beat competing annealers by two orders of magnitude. The paper names one classical solver and no other annealer, gives no timing comparison in numbers, and offers Figure 8, a histogram of total operating time for a full emulated run, as its timing evidence. The numbers may well exist in an internal benchmark prepared for the award jury; the cited sources provide no reproducible timing comparison, and a claim a reader cannot check belongs in the plan column rather than the evidence column.
For a program office the reading is still constructive. The award jury at Messe Stuttgart, which includes Bernd Jungbluth of Fraunhofer, Manfred Bayer of TU Dortmund, Katrin Kobe of Bosch Quantum Sensing and Joachim Ankerhold of Ulm University, judged the hardware concept worth a prize, and the partners are presenting at Quantum Effects in Stuttgart on October 6. The paper gives a clean formulation of a problem every radar house recognizes, an honest error model, and a latency budget with its assumptions on the table. The next public step that would move the result is the one the partners themselves name: the same two scenarios run on the physical C12 processor with an active reset, with parallel readout and timing boundaries documented. Such hardware timing would test the 50 millisecond estimate, move the rung to 4, and allow the question of scale to be asked with real hardware in the loop.
Sources
Primary source: Quentin Schaeverbeke and Viktor Radović (C12 Quantum Electronics, Paris) and Jean-Marc Divanon and Bing Hong Teh (Thales Research and Technology, Singapore), "Simulation of quantum annealing on a semiconducting cQED device for Multiple Hypothesis Tracking (MHT) benchmark," arXiv:2604.15213, posted April 16, 2026. Other material: the C12 and Thales QuantumTrack release of September 29, 2026 as carried by The Quantum Insider; C12's case-study page on the Callisto emulator and radar tracking optimization; Quantum Computing Report, September 29, 2026, on the Messe Stuttgart Quantum Effects Award 2026 winners.