Quandela Quantique Inc. announced MerLin, a new open source framework designed as a detection engine for photonic and hybrid quantum machine learning. Available from February 11, 2026, MerLin integrates optimized quantum simulations into standard machine learning workflows, enabling training and systematic benchmarking of quantum layers. As an initial demonstration, the framework successfully reproduced 18 state-of-the-art photonic and hybrid QML models spanning diverse architectures such as kernel methods and convolutional networks. By embedding photonic quantum models within an established machine learning ecosystem, MerLin allows practitioners to leverage existing tools for comparison and hybrid workflows, and “establishes a shared experimental baseline consistent with widely adopted empirical benchmarking methods in modern artificial intelligence.” This positions MerLin as a tool for linking algorithms, benchmarks, and future quantum hardware.
Advantages of photonic quantum computing in machine learning
Optical quantum computing has proven particularly promising due to its scalability, robustness, compatibility with optical communication technologies, and energy efficiency. This convergence of quantum computing and machine learning offers the potential to extend the capabilities of classical algorithms with quantum machine learning (QML), accelerating progress in both fields. Unlike many approaches, photonic QML “takes advantage of the bosonic nature of light and high-dimensional multimode interference to directly implement and train machine learning models on this unconventional photonic quantum computational model, enabling intrinsic parallelism and efficient exploration of large Hilbert spaces.”
Realizing this potential will require a software framework that bridges abstract QML models and execution on emerging quantum hardware. The need for such tools is emphasized by the current fragmented software environment, where frameworks such as Qiskit, Cirq, Pulser, Perceval, Strawberry Fields, Piquasso, PennyLane, TorchQuantum, and DeepQuantum each specialize in a particular layer or paradigm, hindering algorithm portability. Quandela Quantique Inc. A systematic study reveals that photonic contributions currently account for approximately 6% of all QML publications, but the source of this study is not specified. MerLin distinguishes itself by offering hardware-enabled capabilities that enable testing on existing quantum hardware while enabling exploration beyond current limitations, positioning it as a “future-proof co-design tool that links algorithms, benchmarks, and hardware.” The framework’s design prioritizes systematic benchmarking and reproducibility, which are important needs within the QML community. The researchers identified factors contributing to this need, including “extreme heterogeneity in data preprocessing and task formulation in the literature, and a preference for single-run metrics over multidimensional evaluations.”
To address these challenges, developers have created a replication framework that allows publicly available QML works to be systematically replicated, including performance metrics and experimental analysis. As a demonstration, they successfully reproduced 18 state-of-the-art photonic and hybrid QML studies, validating MerLin’s capabilities and providing a foundation for future development.
MerLin framework integrates PyTorch and Scikit-learn workflows
Currently, the pursuit of practical quantum machine learning (QML) is hampered by a fragmented software environment that requires significant effort to port algorithms across platforms. Frameworks such as Qiskit and Cirq support superconducting processors, and frameworks such as Pulser and Perceval target a variety of hardware, but a unified solution has been lacking. This specialization creates silos and prevents cross-platform experimentation and reproducible research. This is an important question that reflects the challenges faced in the early days of classical artificial intelligence. Recent advances, including DeepQuantum, attempt to fill these gaps, but a comprehensive, integrated approach has remained elusive until now.
Quandela Quantique Inc. is addressing this need with MerLin, a framework designed to embed photonic quantum models directly within established machine learning ecosystems. This integration will allow researchers to “leverage existing tools for ablation studies, cross-modality comparisons, and hybrid classical-quantum workflows.” MerLin excels by combining optimized simulations of linear optical circuits with standard PyTorch and scikit-learn workflows to enable end-to-end differentiable training of quantum layers. This means quantum components can be seamlessly incorporated into existing machine learning pipelines, facilitating a more fluid and iterative development process.
To combat this, the team developed a purpose-built replication framework to systematically replicate 18 state-of-the-art photonic and hybrid QML works. This effort addresses the critical need for “rigorous benchmarking in the QML community,” providing a shared experimental baseline and validating MerLin’s capabilities.
In this study, we introduced MerLin, a quantum software platform designed for large-scale simulation-based exploration of hybrid quantum-classical models while being explicitly hardware-aware.
Reproducing cutting-edge photonic QML experiments
Researchers at Quandela Quantique Inc. and its collaborators are tackling a critical bottleneck in quantum machine learning: reproducibility. This effort addresses a critical problem within the field where fragmented software environments impede independent validation of results. This enables end-to-end differentiable training of quantum layers, an important step towards building practical QML models. This “future-proof co-design tool” aims to link algorithms, benchmarks, and hardware development in a consistent way. These reproductions span a wide range of techniques, including kernel methods, reservoir computing, and convolutional architectures.
They are not simply reimplemented separately. These will be released as reusable, modular experiments aimed at establishing a shared experimental baseline. The team emphasizes that the goal is more than just achieving good results. “Importantly, the goal of this benchmarking effort is not only to identify positive performance improvements, but also to understand their origins,” they explain, emphasizing the importance of disentangling improvements that arise from data preprocessing and model engineering, and those that are truly attributable to quantum mechanisms. The framework and replicated papers are publicly available to foster collaboration and accelerate progress in the field.
Benchmarking challenges and the need for reproducibility
The researchers find that advances in photonic QML “increasingly rely on scalable benchmark-driven experiments rather than individual algorithmic proposals, reflecting the empirical paradigm that underpins modern AI.” This need for systematic evaluation is currently being addressed by new tools designed to establish a shared experimental baseline. Current quantum software frameworks present a fragmented landscape, each specialized in a particular layer or paradigm. Existing options such as Qiskit and Cirq primarily address superconducting processors, while other options focus on specific platforms such as neutral atoms or continuous variable systems. A systematic study found that photonic contributions account for approximately 6% of all QML publications.
To address these issues, a focus on reproducibility is paramount. The goal is not simply to demonstrate performance improvements, but to understand their origins and “separate improvements resulting from data preprocessing, model engineering, or optimization strategies from those resulting from truly new representation or computational mechanisms.” The new framework aims to provide an integrated hardware-aware platform for the implementation, training, and evaluation of photonic QML models, as well as a dedicated replication framework designed for the systematic replication of published works. This includes the reproduction of reported claims, performance metrics, and experimental analyzes within a controlled software environment.
Through an open design and simplified replication framework, MerLin encourages the community to systematically benchmark new and existing results, study learning dynamics, and lower barriers to entry for both classical and QML practitioners.
Various QML software frameworks and their limitations
The proliferation of quantum machine learning (QML) software offers a paradoxical situation: while offering exciting possibilities, the landscape is becoming increasingly fragmented. Many assume a unified ecosystem exists, but in reality, developers are faced with a patchwork of specialized tools, hindering advances in portability and collaboration. Continuous variable approaches are supported by Strawberry Fields and Piquasso, further diversifying your options. However, according to the new framework’s developers, this specialization creates “silos where algorithms cannot be ported without significant transformation work.” Beyond these hardware-specific options, tools like PennyLane, TorchQuantum, and Qiskit-Torch-Module focus on differentiable quantum programming and PyTorch integration.
Recently, DeepQuantum emerged as an integrated platform to bridge qubit circuits, photonic quantum modes, and measurement-based quantum computing, and reported GPU-accelerated gradient computations an order of magnitude faster than PennyLane at large scale. Despite such abundant resources, consistent and adaptable systems remain elusive. The lack of standardization is particularly acute in photonic QML, a promising field that exploits the unique properties of light. Although researchers are increasingly looking at algorithms tailored to specific hardware, “there are currently no existing frameworks that combine efficient simulation, integration with ML workflows, noise models, and hardware access.” This flaw prevents rigorous benchmarking, a key need highlighted by the QML community.
Hardware-aware features and collaborative design possibilities
We are working on a novel framework. MerLin is differentiated by its “hardware-enabled capabilities,” which enable testing on current quantum hardware while facilitating exploration beyond current limitations. MerLin aims to bridge these gaps by providing an integrated platform for photonic quantum machine learning.
