Quandela today announced MerLin, a new quantum computing framework designed to integrate quantum algorithms directly into existing artificial intelligence workflows. Unlike tools aimed at quantum physicists, MerLin empowers data scientists to prototype and benchmark hybrid quantum-classical models using familiar programming environments like PyTorch and scikit-learn. The framework leverages NVIDIA’s CUDA-Q to deliver GPU-accelerated simulation of photonic quantum circuits, enabling research beyond the limitations of current quantum hardware and establishing reproducible metrics for algorithm performance.
MerLin’s Launch and Core Philosophy
MerLin launched at GTC Paris as a framework designed for AI data scientists, contrasting with tools primarily aimed at quantum physicists. Its core philosophy centres on abstracting quantum complexity, enabling practitioners to prototype hybrid quantum-classical models using familiar workflows integrated with PyTorch and scikit-learn.
MerLin prioritises GPU-accelerated simulation of photonic quantum circuits, leveraging NVIDIA CUDA-Q to facilitate testing of algorithms on systems with 24 or more qubits – exceeding the capabilities of current hardware. A key tenet of its design is the establishment of reproducible metrics for hybrid algorithms, addressing a recognised gap in quantum machine learning research where comparative analysis is often hindered by a lack of standardised benchmarks. Integration with Quandela Cloud allows for immediate validation of algorithms on real photonic hardware, enabling study of noise impact and scalability. The framework targets pragmatic applications including quantum-enhanced k-nearest neighbours (kNN), generative adversarial networks (GANs), and variational algorithms, supported by hardware-aware compilation.
MerLin is designed for current photonic quantum processing units (QPUs), specifically those based on Perceval, but its architecture adapts to future hardware generations through features like dynamic circuit recompilation.
Bridging Quantum and Classical Workflows
MerLin facilitates integration between quantum and classical workflows by abstracting quantum complexity into familiar environments. It empowers data scientists to prototype hybrid quantum-classical models utilising existing tools such as PyTorch and scikit-learn, reducing development time. This is achieved through GPU-optimized simulators leveraging NVIDIA CUDA-Q, enabling the testing of algorithms designed for systems with 24 or more qubits – exceeding the capabilities of current quantum hardware.
MerLin establishes reproducible metrics for hybrid algorithms, addressing a noted gap in quantum machine learning research where standardized comparisons are often absent. Integration with Quandela Cloud allows for immediate validation of GPU-optimized algorithms on real photonic hardware, enabling analysis of noise impact and scalability. This supports pragmatic applications including quantum-enhanced k-Nearest Neighbours (kNN), Generative Adversarial Networks (GANs), and variational algorithms, with hardware-aware compilation ensuring efficient execution. The platform’s architecture is designed for current photonic Quantum Processing Units (QPUs), such as those based on Perceval, but is adaptable to future hardware generations through features like dynamic circuit recompilation, ensuring code scalability.
Key Technological Innovations
MerLin incorporates several key technological innovations designed to accelerate quantum machine learning research and development. The framework utilises GPU-optimized simulators leveraging NVIDIA CUDA-Q, enabling testing of algorithms on systems with 24 or more qubits. It establishes reproducible metrics for hybrid algorithms, addressing a recognised gap in quantum machine learning research.
Integration with Quandela Cloud allows for immediate validation of algorithms on real photonic hardware, supporting pragmatic applications including quantum-enhanced k-Nearest Neighbours (kNN), Generative Adversarial Networks (GANs), and variational algorithms, with hardware-aware compilation ensuring efficient execution. MerLin’s architecture adapts to future hardware generations through features like dynamic circuit recompilation, ensuring code scalability.
Early Adoption and User Experience
Early adopters are actively collaborating with Quandela to integrate MerLin into existing workflows and validate its capabilities. These groups include teams from the Perceval Quest, and researchers at Mila, NYUAD’s QML Lab, and Scaleway. Dr. Louis Chen, a Research Associate at the Quantum Centre of Imperial College London (Imperial QuEST) and participant in the Perceval Quest, reports that MerLin facilitated rapid adaptation of existing algorithms to a photonic-native format, yielding comparative insights relevant to ongoing research and publication efforts.
The platform prioritises ease of integration for AI/ML practitioners, enabling prototyping of quantum layers without requiring substantial code rewriting. This is achieved through planned stable APIs for PyTorch and scikit-learn, scheduled for release in Q2 2025. MerLin’s design also supports quantum researchers by providing access to photonic-specific tools, such as boson sampling, alongside GPU-accelerated simulation capabilities. Enterprises can utilise the platform to pilot hybrid quantum-AI workflows, with a focus on establishing clear return on investment benchmarks.
“Quantum shouldn’t demand a PhD to use,” said Niccolo Somaschi, co-founder & CEO of Quandela. “MerLin gives data scientists a GPU-accelerated gateway to quantum advantage while ensuring their work remains compatible with real hardware today—and tomorrow. By integrating benchmarks and noise-aware validation, we’re addressing a critical gap: the lack of reproducible metrics in hybrid algorithm research.”
Future Development and Accessibility
MerLin will be released as a freely accessible resource to encourage widespread adoption, with tiered enterprise options providing access to advanced functionalities. The development roadmap prioritises stable application programming interfaces (APIs) for PyTorch and scikit-learn by Q2 2025, facilitating seamless integration with existing machine learning workflows.
Beyond this, development will focus on expanding support to accommodate photonic quantum processing units (QPUs) with 24 or more qubits, projected for 2026 onwards. This scaling aims to maintain compatibility with evolving hardware capabilities and address the demands of increasingly complex quantum algorithms. The framework’s architecture incorporates dynamic circuit recompilation, a feature designed to ensure code scalability and adaptability across successive generations of quantum hardware, mitigating the risk of code obsolescence as quantum technology advances.
More information
External Link: Click Here For More
