Tide Surge AI Editorial Team

On September 24, Tsinghua-affiliated quantum AI company Sober Heterogeneous announced the completion of its A+ round of financing, with investors including Jingkai Capital, Anhui High-Tech Investment, Daode Investment, Lishi Investment, China Consulting Fund, Xuhui Capital, Senlan Group, and the founder of Jingdong Group.

Combined with its previous Series A round, the company has secured two rounds of financing within three months, totaling nearly 100 million yuan.

Founded in 2021 by Yu Teng, an alumnus of Tsinghua University's Department of Computer Science, Sober Heterogeneous initially focused on parallel heterogeneous computing and automatic parallel compilation, completing three consecutive financing rounds within six months in 2022.

In 2025, the team completed mathematical validation of a high-density intelligent construction method based on quantum computing principles. Yu Teng subsequently halted traditional AI operations, shifted R&D focus to quantum AI, and established operations in Shanghai's Xuhui District.

This June, Sober Heterogeneous released RiverONE, China's first visual language model reconstructed using simulated quantum computing power.

The model uses simulated quantum computing to generate parameters during construction, then performs inference on classical GPUs after training, without requiring real-time quantum computer participation.

According to third-party test reports, RiverONE, with 1.9 billion parameters, achieved at least 95% of the performance of NVIDIA's 35 billion-parameter Ising Calibration 1 in quantum calibration chart understanding tests.

The product has integrated with multiple chip and computing power vendors.

Sober Heterogeneous became an AMD AI system supplier in 2023; Biren Technology completed compatibility testing for mainstream quantum computing frameworks including Cirq, PennyLane, and Qiskit on its GPUs, with no abnormal interruptions during 48-hour stability tests; Moore Threads' Xieyun C-series GPUs completed RiverONE adaptation for the vLLM inference framework; and Chuangxin Electronics collaborated with Loongson to achieve Day0 compatibility for Sober Heterogeneous's full suite of quantum-inspired models and operators.

In terms of commercialization, after domestic leading optical lens manufacturer Senlan Optics introduced Sober Heterogeneous's quantum optimization model, precision processing of optical raw materials shifted from iterative trial-and-error to one-time completion after scanning, improving overall efficiency by 20% to 30%.

Yu Teng estimates that each high-precision optical processing equipment thus generates over 1 million yuan in additional value.

In August this year, Sober Heterogeneous signed a strategic cooperation agreement with quantum computing company Taiyi Quantum to jointly develop visual language models for neutral atom quantum computer debugging.

Sober Heterogeneous's completion of two financing rounds within three months was driven by the successful implementation of quantum-generated parameter methodologies on classical GPUs.

This approach bypasses the bottleneck of immature quantum hardware, enabling quantum-inspired models to be directly deployed on existing computing infrastructure.

Compatibility records with AMD, Biren, Moore Threads, Chuangxin, and Loongson serve as intermediate evidence of this pathway's transition from laboratory to production lines.

However, validation remains concentrated in two narrow scenarios—quantum calibration and optical processing. When expanding to broader industrial applications, engineering delivery capabilities become more critical than model metrics.