In recent years, the convergence of quantum computing, artificial intelligence (AI), and high-performance computing (HPC) has become a central theme in the evolution of scientific computing infrastructure. From AI4Science and quantum machine learning to supercomputing centers and heterogeneous computing platforms, phrases such as "Quantum x AI x HPC", "integrated quantum-supercomputing-intelligence infrastructure", and "next-generation research infrastructure" now appear frequently in academic conferences, industry forums, and corporate presentations.
At the same time, a clear pattern has emerged:
The messaging is becoming increasingly similar, while the actual technical depth of different products varies dramatically.
Whether the subject is a quantum software platform, an AI4Science infrastructure stack, or a heterogeneous computing framework, many projects now describe themselves in similar terms:
- integrating quantum computing with AI;
- supporting heterogeneous computing resources;
- serving as future research infrastructure;
- enabling applications in materials science, chemistry, biomedicine, and other industries;
- building an open ecosystem and developer community.
These directions are meaningful. In fact, they are becoming part of the field's shared consensus.
The real question is different:
When everyone is telling a similar story, how can we tell whether a platform has actually delivered technical substance, rather than remaining at the level of conceptual packaging and slideware?
For scientific infrastructure, it is more useful to ask five verifiable questions than to focus on slogans:
- Is it open source?
- Does it provide public benchmarks?
- Is it used continuously by high-quality research communities?
- Has it supported real industry-oriented application cases?
- Does it continue to evolve through sustained version updates?
Any platform that claims to be "next-generation research infrastructure" should be able to answer these questions in a concrete way.
The development of TensorCircuit-NG offers a useful case study. Its value does not lie only in proposing a vision for "Quantum x AI x HPC"; it lies in a body of work that can be inspected, reproduced, cited, extended, and tested over time: open code, reproducible performance evaluations, a visible record of academic adoption, evidence of industry spillover, and six years of engineering iteration.
1. Is It Open Source?
For scientific software, open source means more than publishing code.
It means that:
- the technology can be independently verified;
- performance claims can be reproduced;
- algorithms can be inspected;
- users can deploy the software without relying on a closed service;
- third-party researchers can repeat experiments under their own conditions.
Research communities do not lack polished presentations. What is much rarer is a technical system that can survive independent inspection.
TensorCircuit was not released as a one-off code dump. Its development forms a traceable engineering trajectory: from the original personal open-source version, to the version developed during the Tencent Quantum Lab period, and then to the currently maintained TensorCircuit-NG project. Across these stages, the core code, documentation, tests, and examples have remained open. The GitHub history preserves the development record, with more than 500 combined stars and forks, over 2,700 commits, more than 30 released versions, and contributions from over 30 developers around the world.
In terms of engineering scale, TensorCircuit-NG is no longer a short-term proof-of-concept project. It is a platform-level scientific computing system, with roughly 70,000 lines of code, type annotations, unit tests, continuous integration, documentation, and tutorials. The repository currently contains close to one thousand test functions. These tests are not merely a coverage metric; they are part of the engineering foundation that keeps APIs stable, backend behavior consistent, and long-term maintenance manageable.
The surrounding ecosystem matters as well. TensorCircuit-NG provides documentation, more than 30 tutorial examples, over 170 application examples, more than 10 benchmark suites, and a companion quantum computing tutorial. Together, these resources form a developer ecosystem that is learnable, reusable, and extensible. The platform also embraces AI-native workflows by providing AI skill packages for paper reproduction, code translation, and performance optimization. This means TensorCircuit-NG is not only designed for human developers; it is also adapting to a new mode of scientific software development in which AI agents participate directly in research workflows.
Another measurable signal of open-source adoption is installation and use. TensorCircuit-related packages include tensorcircuit, tensorcircuit-ng, and the nightly package tensorcircuit-nightly on PyPI, with cumulative pip install downloads exceeding one million. Download counts alone do not prove scientific value, but they do show that the platform exists in real development environments, not only in papers or promotional pages.
For research infrastructure, credibility comes from the ability of third-party users to run the code, inspect the implementation, reproduce experiments, and build their own workflows. Code is always more honest than marketing material.
2. Are There Public Benchmarks?
Every computing platform eventually has to answer a simple question:
Does it actually improve computational efficiency?
This is why public benchmarking is essential for judging platform maturity. In a high-performance setting such as "Quantum x AI x HPC", claims about acceleration, heterogeneous execution, or scalability are difficult to evaluate without reproducible benchmarks.
One of the earliest reasons TensorCircuit attracted attention was its benchmark system for differentiable quantum computing and tensor-network simulation. The first TensorCircuit white paper was published in Quantum: TensorCircuit: a Quantum Software Framework for the NISQ Era. The paper introduced the platform architecture, core functionality, and performance advantages, and compared TensorCircuit against several mainstream quantum software frameworks on variational quantum algorithms, gradient computation, and quantum circuit simulation.
The work made clear why unified tensor programming, automatic differentiation, and just-in-time compilation matter for quantum computing workflows. In several variational quantum algorithm and gradient computation tasks, TensorCircuit demonstrated significant performance advantages over representative frameworks such as IBM's Qiskit and PennyLane, with speedups reaching multiple orders of magnitude in some cases. More importantly, these results were not confined to figures in a paper: the code, experimental setup, and evaluation procedures were made reproducible. That is the difference between a verifiable technical path and an unverifiable performance claim.
With the release of TensorCircuit-NG, the benchmark scope has expanded toward problems closer to future research infrastructure:
- GPU-accelerated computing;
- optimized tensor-network contraction;
- distributed HPC environments;
- unified computation graphs spanning quantum circuits, neural networks, and tensor networks.
The NG white paper further summarizes TensorCircuit's upgrade toward the integration of quantum computing, supercomputing, and intelligent computing; see the preprint. The focus has shifted from "how to simulate quantum circuits faster on a single machine" to "how to organize quantum, AI, and numerical computing workflows in realistic heterogeneous research environments."
External evaluation provides another layer of evidence. NVIDIA used TensorCircuit as a third-party quantum software case in its cuQuantum 23.10 benchmarking context. This shows that TensorCircuit has entered the evaluation landscape of hardware and high-performance computing vendors. For scientific infrastructure, such external benchmarks complement open papers and are more persuasive than slide-based claims.
3. Is It Used by the Research Community?
For scientific infrastructure, the hardest signal to fake is not performance.
It is sustained use by serious research communities.
A platform can gain short-term attention through marketing, but it cannot gain long-term citations through marketing alone. Research adoption is a form of long-horizon voting. If a platform continues to support high-quality work across institutions, research areas, and teams, then it has demonstrated real utility.
More than 170 academic works have cited TensorCircuit, and in the first five months of 2026 alone, more than 40 works have already cited it. More importantly, these works are not concentrated in a single niche. They span quantum simulation, quantum machine learning, quantum chemistry, quantum sensing, quantum architecture search, and AI4Science.
Quantum Simulation and Many-Body Systems
In many-body quantum physics, condensed matter systems, and complex quantum dynamics, researchers often need large-scale quantum circuit simulation, tensor-network contraction, and differentiable optimization. These tasks place high demands on performance, numerical stability, and automatic differentiation.
Representative works include Zero and Finite Temperature Quantum Simulations Powered by Quantum Magic from teams including NVIDIA, Google, MIT, and Harvard; Exploring nontrivial topology at quantum criticality in a superconducting processor from Haohua Wang's group at Zhejiang University; and Variational LOCC-assisted quantum circuits for long-range entangled states from Xiongfeng Ma's group at Tsinghua University. These papers show that TensorCircuit is not limited to abstract algorithm demonstrations; it is being used in concrete problems in many-body physics and experimental quantum information.
Quantum Machine Learning
Quantum machine learning is one of the most active application areas for TensorCircuit. Representative papers include Understanding quantum machine learning also requires rethinking generalization from Jens Eisert's group at the Free University of Berlin, Dynamical transition in controllable quantum neural networks with large depth from teams including Liang Jiang and Junyu Liu, Generative Quantum Machine Learning via Denoising Diffusion Probabilistic Models from Quntao Zhuang's group at the University of Southern California, and IBM Quantum's Dynamic parameterized quantum circuits: expressive and barren-plateau free.
These works all require stable workflows connecting parameterized quantum circuits, gradient computation, model training, and numerical simulation. TensorCircuit's value is visible precisely at this workflow level: it connects quantum circuit simulation, automatic differentiation, and machine learning training into a unified programmable system.
Quantum Architecture Search and Algorithm Design
TensorCircuit has also been used in algorithmic and learning-theoretic research. Examples include Learning Quantum States and Unitaries of Bounded Gate Complexity from Caltech and Google, Quantum Machine Learning Architecture Search via Deep Reinforcement Learning from Brookhaven National Laboratory, and Distributed quantum architecture search from Luzhou Li's group at Sun Yat-sen University.
This class of work highlights the platform's infrastructure role. Researchers are not merely calling a fixed algorithm; they are building new search strategies, learning processes, and experimental protocols on top of TensorCircuit.
Quantum Chemistry and Fermionic Simulation
The quantum chemistry ecosystem around TenCirChem further extends TensorCircuit's application boundary. Quantum chemistry and fermionic simulation typically require complex Hamiltonian construction, differentiable optimization, tensor-network representations, and high-performance simulation. They therefore provide a demanding test case for any scientific computing platform.
Representative works include Efficient quantum simulation of electron-phonon systems by variational basis state encoder from teams at Tsinghua University and The Chinese University of Hong Kong, Shenzhen, as well as Fast Emulation of Fermionic Circuits with Matrix Product States from Garnet Chan's group at Caltech. These studies show that the TensorCircuit ecosystem has moved from general quantum circuit simulation into more specialized domains such as quantum chemistry.
Quantum Sensing and Imaging
TensorCircuit has also been used in quantum sensing, imaging, and experiment-facing tasks. Examples include End-to-end variational quantum sensing from Roger Melko's group at the Perimeter Institute, and Practical advantage of quantum machine learning in ghost imaging from Guihua Zeng's group at Shanghai Jiao Tong University. These works illustrate the platform's potential in quantum sensing and measurement-related applications.
The value of a research platform is not captured by a single paper. It is reflected in its ability to support many research directions over time. More than 170 citing works, users across high-level institutions, and multiple examples in leading journals and conferences form an evidence chain that is stronger than any single promotional claim.
4. Has It Supported Industry-Oriented Applications?
Academic citations show whether a platform can support research. Industry-oriented application cases show whether it can move toward real-world problems.
It is important to be precise here. Quantum computing is still exploratory in many industrial contexts, so the right question is not whether it has already replaced classical solutions at scale. The better question is whether researchers and engineering teams in different fields have used the platform to build prototypes, workflows, and validation pipelines for real problem domains. From this perspective, TensorCircuit's application spillover already reaches multiple sectors.
In agricultural diagnostics, researchers have used a quantum vision transformer for tomato leaf disease detection; see Enhancing Agricultural Diagnostics: Tomato Leaf Disease Detection Using Quantum Vision Transformer. In neuroscience and medical imaging, related works include Predicting Brain Age and Gender from Brain Volume Data Using Variational Quantum Circuits and Expanding the Horizon: Enabling Hybrid Quantum Transfer Learning for Long-Tailed Chest X-Ray Classification. In drug discovery, A hybrid quantum computing pipeline for real world drug discovery explores a hybrid quantum computing workflow for real drug discovery problems.
TensorCircuit-NG has also appeared in security, communications, optimization, and computing systems. In software security, researchers have proposed lightweight quantum convolutional neural networks for malicious code detection. In drone and radar applications, hybrid quantum neural networks have been explored for radar return signal processing. In edge computing, quantum reinforcement learning has been used for joint resource allocation and task offloading. In finance, improved QAOA methods based on conditional value-at-risk have been studied for portfolio optimization. The significance of these cases is that they move quantum software frameworks from "quantum algorithm papers" into concrete domains such as agriculture, medicine, security, communications, finance, and drug discovery.
External recognition provides additional context for the ecosystem. TensorCircuit has appeared in PhotonBox's 2022 list of influential quantum industry events in China, was listed as a recommended quantum software project in Google Summer of Code 2023, was used by NVIDIA in cuQuantum evaluation materials, was invited to participate in UnitaryHack 2024, and participated in Open Source Promotion Plan 2025. These forms of recognition do not replace technical validation, but they do show that TensorCircuit is not an isolated lab project. It has entered the public view of the open-source quantum software and high-performance computing ecosystems.
Industrial maturity does not happen overnight. It typically moves from research prototypes, to open tools, to cross-domain collaboration, to engineering validation, and eventually to deployment. TensorCircuit-NG's current value lies in providing a reusable low-level toolchain for that process.
5. Does It Continue to Evolve?
One defining feature of scientific infrastructure is that it is never finished.
New hardware appears. New algorithms appear. New scientific demands appear. This makes sustained iteration more important than a single innovation.
TensorCircuit's history is a good example. The project was first released in April 2020. From 2020 to 2021, TensorCircuit completed its core architecture, automatic differentiation mechanism, and early quantum algorithm modules, establishing the academic foundation for a unified tensor-computing framework. From 2021 to 2024, under the Apache License 2.0, the project continued to evolve in engineering: performance optimization, interface standardization, multi-backend support, and community ecosystem development gradually turned it into an open-source platform for global research users and developers.
Since the launch of TensorCircuit-NG, or "Next Generation", in 2024, the project has moved beyond a quantum computing software framework toward a broader next-generation research infrastructure. It explores deeper integration among quantum computing, supercomputing, and intelligent computing, while continuing to expand its ecosystem in AI4Science and related areas.
Sustained iteration is also visible in upstream and downstream ecosystem contributions. Upstream, core developers have contributed to standard machine learning frameworks such as TensorFlow, including work related to the automatic differentiation formula for complex-valued singular value decomposition and fixes to vectorized matrix multiplication. In the tensor-network ecosystem, TensorNetwork-NG continues to maintain the original Google TensorNetwork framework and keep it usable. Downstream, TenCirChem extends TensorCircuit capabilities into quantum computational chemistry workflows.
These upstream and downstream contributions show that TensorCircuit-NG does not confine itself to a single framework. Instead, it builds connections among machine learning, tensor networks, quantum chemistry, and high-performance computing. This matters for Quantum x AI x HPC integration, because future research infrastructure cannot serve only one model family, one hardware type, or one class of algorithms.
In the TC-NG architecture:
- quantum circuits;
- neural networks;
- tensor networks;
are brought into a unified computation-graph system.
At the same time:
- CPUs;
- GPUs;
- HPC clusters;
- QPUs;
are becoming part of a unified resource pool.
This marks a shift in platform positioning: from a quantum software framework to infrastructure for future scientific computing. Compared with projects that remain at the stage of concept demonstrations, short-term packaging, or slide-based roadmaps, more than six years of open-source development, continuous iteration, and repeated research-community validation say much more about a platform's real engineering capacity and long-term value.
Conclusion: What Builds Trust in Scientific Infrastructure?
In the rapid development of Quantum x AI x HPC, industry narratives are converging.
More and more platforms now talk about:
- AI4Science;
- hybrid quantum-classical computing;
- scientific research infrastructure.
These directions are worth pursuing. But for users, researchers, and industry partners, the core criteria have not changed:
Is the platform fully open source?
Does it provide public benchmarks?
Is it broadly and continuously used by high-quality research communities?
Has it supported cross-industry application cases?
Does it continue to evolve through sustained version updates?
Once these questions are answered one by one, the value of a platform does not need to depend on slogans or conceptual messaging. For scientific infrastructure, long-term trust is built on verifiable code, reproducible experiments, growing academic adoption, application spillover into real problems, and engineering iteration that stands the test of time. In an era where technical narratives increasingly sound alike, these qualities are especially valuable.
Top comments (0)