Fermi Universe launches quantum-enhanced LLM FermiQLLM 1.0 on a 100 million yuan seed round

A Chinese startup called Fermi Universe has surfaced with a pitch its founders describe as the country's first company dedicated to Quantum for AI. QbitAI reports that the company has released FermiQLLM 1.0, which it calls the world's first full-chain quantum-enhanced large language model.

On funding, Fermi Universe is at seed stage with a cumulative 100 million yuan raised, at a post-money valuation of roughly 1 billion yuan, with capital still being transferred. QbitAI describes the round as the largest seed financing in China's Q4AI field and a key reason venture investors have taken notice. The report is presented as exclusive.

The team is heavy on Tsinghua ties. Nearly half the staff come from the university, mainly from its science foundations programme, its computer science AI lab and its high-performance computing lab, and a quantum research group at Tsinghua's Yang Zhenning Institute for Advanced Study is collaborating on projects. Founder and CEO Li Wei was Hunan province's top science scorer in the 1998 college entrance exam, a Tsinghua computer science NLP lab alumnus of the class of 1998, a former executive at NetEase Youdao and Renren, and the 2005 translator of a foundational Chinese text on statistical natural language processing.

The physics matters more than the branding. Fermi Universe's quantum enhancement does not mean running a full large model on quantum hardware; instead the team embeds quantum physics and quantum many-body methods such as tensor networks, quantum simulated annealing and gauge degrees of freedom into the classical pipeline that carries a model from data through training to evaluation. Internal sources say the approach covers five stages, from data representation through reinforcement to evaluation, and does not depend on fault-tolerant quantum computers, so training and deployment can run on existing GPU infrastructure.

FermiQLLM 1.0 is built by re-engineering an open-source Qwen base model. In internal tests, the company says inference performance rose by more than 15% against traditional models of comparable parameter count, sustained reinforcement learning training cost fell by more than 25%, and composite scores on MATH-500, GPQA-Diamond and BBH improved by 10% to 20% over same-scale open base models.

The timeline, per QbitAI: the project began in March this year with experiments on a 4B model to prove feasibility; the company was formally founded in May, two months later; larger-scale quantum-enhanced development followed, and the first product has arrived roughly in step with the seed round.

Yao Hong, deputy director of Tsinghua's Yang Zhenning Institute for Advanced Study and a fellow of the American Physical Society, told QbitAI that realising quantum physics's contribution to large models will require systematically folding ideas such as entanglement, correlation effects and topological physics into architecture design, training and inference. In his view Quantum for AI is still at an early industrial stage, and industry-academia work of this kind could shape the next generation of AI.

The wider field splits roughly two ways. One camp leans on quantum hardware, using large models as agents that call quantum devices for specific optimisation jobs; the other sidesteps hardware maturity by importing quantum algorithms and mathematics into classical computing, which is the path Fermi Universe has taken. Overseas players include SandboxAQ, focused on post-quantum cryptography and quantum-inspired optimisation, Microsoft's Azure Quantum Elements, aimed at AI for science and simulation, and Xanadu, which offers a quantum machine learning research framework. Which layer quantum should help AI at remains unresolved, at home and abroad.

Watch next: over the coming 6 to 12 months the company plans to abstract its single-base-model work into a general framework that adapts to different models with automated iteration, and to open up module evaluations and open-source part of the technology. Whether those promises hold, and whether outsiders can reproduce the internal numbers, is the real test of the approach.