The rise of quantum computing is redefining the limits of machine learning, but most current quantum kernel methods suffer from severe practical limitations: they scale quadratically with training set size, use fixed non-trainable kernels, and lack an intrinsic formulation for multiclass classification. Against this backdrop, McQuack emerges as a trainable quantum kernel that promises to change the game. In this article we explore in depth what McQuack is, how it works, what results it has achieved in simulation and on real IBM hardware, and how this technology can be integrated into today’s business ecosystem, hand in hand with custom software and artificial intelligence solutions.
Traditional kernels, such as RBF or polynomial, are powerful tools, but their application on large datasets becomes unfeasible: the full Gram matrix requires O(n²) memory and computation. Moreover, once the kernel is chosen, it cannot be adapted during training, limiting its ability to fit complex patterns. Finally, multiclass classification forces the use of one-vs-one or one-vs-all schemes, adding overhead and complexity. McQuack tackles all three fronts simultaneously.
The key innovation of McQuack lies in replacing the expensive full Gram matrix with a fidelity matrix between each sample and class centroids. Instead of comparing every pair of samples, the model computes the quantum fidelity between a sample and a set of trainable centroids. This reduces complexity from O(n²) to O(n × k), where k is the number of classes, achieving linear scaling in the number of training samples. Furthermore, the centroids are not fixed: they are optimized during training using variational algorithms, making the kernel adaptive. This trainable capability allows the model to learn which regions of the Hilbert space are relevant for each class, improving separation without manual feature engineering.
But the most groundbreaking aspect is that McQuack handles multiclass classification naturally. By defining one centroid per class, prediction is made by assigning the sample to the centroid with highest fidelity, without the need for binary decompositions. This simplifies the model, reduces computational cost, and provides a direct geometric interpretation: each class occupies a region around its centroid in the quantum state space.
The experiments conducted by the McQuack team cover over 150 datasets, both in simulation and on real IBM hardware with up to 124 qubits. In simulation, McQuack outperformed existing “pure” quantum methods, such as traditional variational classifiers, demonstrating higher accuracy and efficiency. On real hardware, without prior training (direct inference), results were comparable to a classic RBF kernel, suggesting that even without optimization the model inherits useful properties from the Hilbert space.
A critical aspect for the practical viability of any trainable quantum model is the presence of barren plateaus that hinder optimization. McQuack found no evidence of barren plateaus in experiments with up to 13 qubits, and the researchers emphasize that parameter initialization is crucial for success. This opens the door to scaling to larger circuits without fear of gradient vanishing.
From a business perspective, McQuack represents a significant advance for those seeking to integrate quantum capabilities into their machine learning workflows, without having to deal with the limitations of classical kernels. This is where expertise in AI and custom software development becomes indispensable. At Q2BSTUDIO, we accompany companies in adopting these disruptive technologies, designing solutions that integrate quantum models with robust cloud infrastructures. Whether deploying a McQuack classifier on AWS or Azure, or combining it with Business Intelligence systems like Power BI for real-time analytics, the key is to build flexible ecosystems that maximize return on investment.
Furthermore, the trainable nature of McQuack fits perfectly with modern autonomous AI agent approaches. An intelligent agent could use McQuack as an online classification module, adapting its centroids as it receives new data, without needing to retrain from scratch. This enables continuous learning systems, ideal for dynamic environments such as cybersecurity, where attack patterns constantly evolve. At Q2BSTUDIO, we develop AI agents capable of integrating these quantum kernels to detect anomalies and prevent intrusions, all supported by a secure and scalable cloud architecture.
Of course, cybersecurity is a fundamental pillar in any quantum technology implementation. Quantum models, like classical ones, are vulnerable to adversarial attacks and require protection at all layers. Our cybersecurity services include pentesting and vulnerability analysis to ensure that McQuack deployments in cloud environments are resilient. Likewise, the linear scalability offered by McQuack allows processing large volumes of data without sacrificing performance, which is critical for real-time BI applications where every millisecond counts.
In conclusion, McQuack is not just an academic advance: it is a practical tool that brings quantum computing closer to real-world multiclass classification problems. Its linear scaling, trainable capability, and natural multiclass formulation make it an ideal candidate for integration into enterprise platforms. At Q2BSTUDIO, we combine this technology with our expertise in custom software, cloud AWS/Azure, AI, cybersecurity, BI with Power BI, and AI agents to deliver unique solutions that make a difference. The future of quantum machine learning is already here, and it is trainable.
