Chinese Startups Push Quantum Computing to Desktops and Into Large-Model Training

Unitary Quantum was incubated by a research team from Shanghai Jiao Tong University. At its launch event, the company said it wants users to describe a scientific problem in plain language and let the system run the quantum computation. Running a quantum algorithm has required knowledge of quantum circuits, specialized code, quantum simulators with GPU acceleration or connections to real quantum machines, and decisions about whether data can be uploaded to the cloud. Those barriers have kept most developers out.

UnitarySpark is described as the company's first-generation heterogeneous computing base product. It combines quantum simulation, GPU acceleration, AI Agent interaction and real-machine interconnection in a single workstation. The company says inference, algorithm execution and data storage all happen locally, so experimental parameters and model configurations do not need to go to the cloud. For scenarios such as financial modeling and drug design, keeping data on-premises is a compliance requirement, not just an advantage. UnitarySpark also reserves an interface for real quantum machines. A unified Circuit API connects simulators and real quantum processors, allowing code to move from simulator to real hardware without changes or switching tools. Unitary Quantum has partnered with Guokun Quantum and other quantum hardware companies to verify this simulation-to-real-machine path.

Both UnitarySpark and UnitaryLab include AI Agents. Users can describe a partial differential equation, a type of quantum algorithm or an engineering simulation in natural language. The Agent translates the request into a mathematical problem, algorithm parameters and executable tasks, then schedules the underlying compute and runs the result. A September 2026 NVIDIA website success case cited about 12 manual coding steps being compressed into roughly three natural-language descriptions, cutting interaction frequency by about 75 percent. Zhang Lei, co-founder of Unitary Quantum, said the next bottleneck for quantum computing is not only compute itself but how to let users make good use of it. With an Agent, he said, researchers can describe a problem and the system runs the full process from parameter generation to simulation execution, lowering the barrier to quantum computing.

Hu Junpeng, CTO of Unitary Quantum, described UnitarySpark as a quantum computing workbench that can sit on a desk, not a simple local device. The company's underlying algorithmic asset is an original framework called Schrödingerization, developed by Shanghai Jiao Tong University's quantum scientific computing team. The method uses a mathematical transformation to turn non-unitary differential equation problems into unitary evolution forms that quantum systems can handle directly. It was selected for the National Natural Science Foundation of China's 14th Five-Year Plan outstanding achievements compilation, named the only representative achievement in mathematics in 2024, and included in Shanghai Jiao Tong University's 2030 Plan. Unitary Quantum was founded in 2025 by the university's research team to move quantum computing from the lab to industry. On top of Schrödingerization, the team built UnitaryLab, which links algorithm design, quantum simulation, circuit construction, real-machine execution, intelligent task organization and result verification.

The company said UnitaryLab's work based on NVIDIA DGX Spark was selected as an NVIDIA website success case. It also entered the first AI for Industry capability list and was named a typical case in the 2026 Excellent Domestic Autonomous Industrial Simulation Software program. At the launch event, Unitary Quantum signed strategic cooperation agreements with Shanghai Guokun Quantum, Shanghai Bose Yuanpei Quantum and Shanghai Zhongqi Wuliang Quantum to work on software-hardware adaptation, hardware interconnection and scenario validation. Additional cooperation intentions cover finance, meteorology, biomedicine, artificial intelligence and quantum education.

Fermi Universe is trying to bring quantum ideas into AI model development. QbitAI reported that it is China's first startup focused on Q4AI, or Quantum for AI, and that nearly half of its employees have Tsinghua University backgrounds, mainly from the basic science class, the computer science department's artificial intelligence lab and the high-performance lab. A quantum research team from Tsinghua University's Yangzhenning Institute for Advanced Study is collaborating with the company. The company is in a seed round, with 100 million yuan raised, the largest seed round in China's Q4AI field, at a post-money valuation of about 1 billion yuan. The funds are being delivered, according to QbitAI.

Fermi Universe recently launched FermiQLLM 1.0. Internal sources cited by QbitAI said the model completes systematic quantum enhancement in five stages: data representation, model structure, model training, model reinforcement and model evaluation. It is compatible with mainstream AI compute infrastructure and can be deployed industrially. The approach does not run the full large model on a quantum computer. Instead, it embeds quantum physics and quantum many-body computing methods, including tensor networks, quantum simulated annealing and gauge degrees of freedom, into the complete chain from data to training to evaluation, while remaining within classical computing. The company said FermiQLLM 1.0 does not depend on fault-tolerant quantum computers and can be trained and deployed on existing GPUs and other infrastructure. The project began in March, with a 4B small model used for feasibility tests; the company was founded in May and then began quantum-enhanced research on larger parameter models. FermiQLLM 1.0 is based on the open-source Qwen base model. In internal tests, the company said inference performance improved by more than 15 percent against a traditional model of the same parameter scale, and training cost in the continual reinforcement learning stage fell by more than 25 percent. On benchmarks including MATH-500, GPQA-Diamond and BBH, the model's overall performance improved by 10 percent to 20 percent over a same-scale open-source base model.

Fermi Universe founder and CEO Li Wei was the top science student in Hunan province's 1998 college entrance exam and a 1998 alumnus of Tsinghua University's computer science NLP lab, with bachelor's and master's degrees from Tsinghua. He previously held senior technology roles at NetEase Youdao and Renren, including AI technology committee chair, and in 2005 translated Foundations of Statistical Natural Language Processing. Yao Hong, vice dean of Tsinghua University's Yangzhenning Institute for Advanced Study and an American Physical Society Fellow, told QbitAI that fully using quantum physics to empower large models will require systematically integrating ideas and methods from quantum entanglement, correlation effects and topological physics into model architecture design, training optimization and inference generation. He said Q4AI is still in an early industrial stage and that the collaboration between the institute's quantum research team and Fermi Universe is an industry-level practice combining production, education and research.

The quantum-AI field currently has two main approaches, according to QbitAI. One relies more on quantum hardware, such as using a large model as an Agent to call quantum computing devices for specific optimization tasks. The other avoids the maturity limits of quantum hardware by bringing quantum algorithms, quantum physics and related mathematical tools into classical computing to transform AI models. Fermi Universe takes the second route. Overseas players include SandboxAQ, which focuses on post-quantum cryptography and quantum-inspired optimization; Microsoft Azure Quantum Elements, which leans toward AI for Science and scientific simulation; and Xanadu, which provides research frameworks for quantum machine learning. In China, companies are pursuing both routes, and technical paths have not converged. Over the next six to 12 months, Fermi Universe plans to abstract its quantum enhancement work on a single base model into a general framework that can adapt to different large models and automate iterative optimization. It also plans to gradually open module evaluations and push some technology into open source.

Editor's Summary

Unitary Quantum launched UnitarySpark and UnitaryLab 2.5, aiming to bring local, natural-language-driven quantum computing to a desktop workstation with data kept on-premises. Fermi Universe unveiled FermiQLLM 1.0, a quantum-enhanced large model trained on classical infrastructure, after raising 100 million yuan in a seed round. Both moves show Chinese startups trying to make quantum technology more usable in scientific computing and AI model development.