- Fermi Universe raised 100 million yuan in seed funding at a post-money valuation of about 1 billion yuan, described as the largest seed round in China's Q4AI sector.
- Its FermiQLLM 1.0 model applies quantum-inspired methods such as tensor networks and quantum simulated annealing to classical GPU training, not fault-tolerant quantum hardware.
- Internal benchmarks claim 10-20% improvement over same-scale open-source base models on MATH-500, GPQA-Diamond and BBH.
- The model is built on Alibaba's Qwen open-weight base and targets compatibility with existing AI compute infrastructure.
Seed Funding and Tsinghua Roots
In a QbitAI interview originally published on September 27, 2026, Fermi Universe was described as China's first startup focused on Quantum for AI (Q4AI). The company has accumulated 100 million yuan in seed-stage financing with funds still being disbursed, representing what QbitAI called the largest seed round in the domestic Q4AI field, corresponding to a post-money valuation of approximately 1 billion yuan.
The founding team is heavily drawn from Tsinghua University. Founder and CEO Li Wei graduated from Tsinghua's computer science department and previously held senior AI roles at NetEase Youdao and Renren. Nearly half of the staff come from Tsinghua's basic science program, AI laboratory and high-performance computing laboratory. The company also collaborates with the Yang Zhenning Institute for Advanced Study at Tsinghua and serves as an internship site for the university's Q4AI research direction.
Quantum-Inspired Methods on Classical Hardware
Fermi Universe's approach does not require fault-tolerant quantum computers. Instead, it embeds techniques from quantum physics and quantum many-body computation — including tensor networks, quantum simulated annealing and gauge degrees of freedom — into every stage of the large-model pipeline on conventional GPU clusters. The company calls this 'quantum enhancement' rather than quantum-native execution.
The rationale is that quantum physics studies complex many-body systems in high-dimensional Hilbert spaces, whose superposition, entanglement and strong-correlation effects share structural parallels with the latent representations learned by large language models. By importing these mathematical tools into classical training workflows, the team aims to improve efficiency without waiting for mature quantum hardware.
FermiQLLM 1.0 Performance Claims
The company's first product, FermiQLLM 1.0, is built on Alibaba's Qwen open-weight base model. According to internal testing reported by QbitAI, the quantum-enhanced version delivers more than 15% higher inference performance and reduces reinforcement-learning training costs by over 25% compared with traditional models of the same parameter scale.
On mainstream benchmarks including MATH-500, GPQA-Diamond and BBH, Fermi Universe claims a 10% to 20% comprehensive improvement over same-scale open-source base models. These figures come from the company's own evaluations and have not been independently verified.
Over the next six to twelve months, the team plans to generalize its quantum-enhancement capabilities into a framework compatible with multiple base models and to release some components as open source.
Sources & context
Go to the original material. Company claims remain attributed to their sources.
01
