发生了什么
Qingxing Heterogeneous Computing, a startup incubated at Tsinghua University, has completed a Series A+ funding round, bringing its total capital raised over the last three months to nearly 100 million yuan (approximately $14.9 million). The company is focused on developing 'quantum-inspired' AI models, specifically its RiverONE vision-language model, which utilizes simulated quantum computing techniques during the construction phase to generate parameters that allow the model to run efficiently on standard GPUs.
Qingxing, also known as Sober Heterogeneous, closed its Series A+ round following an earlier Series A financing. Investors in the latest round include Jinkai Capital, Anhui High-Tech Investment, Daode Investment, Lishi Investment, Zhongzi Fund, Xuhui Capital, Senlan Group, and the founder of JD.com Group.
The company's flagship product, RiverONE, is a 1.9-billion-parameter vision-language model. According to a third-party test report cited by PE Daily AI, RiverONE achieved at least 95% of the performance of the 'NVIDIA Ising 1' model on a specialized task involving the interpretation of quantum calibration charts, while using less than one-tenth of the parameters.
Qingxing is actively collaborating with Chinese chipmakers, including Biren Technology, MetaX, and Taichu Electronics, to adapt its software and models to their specific hardware architectures. These efforts include testing on various GPUs and integrating with software frameworks like vLLM to ensure compatibility across different computing platforms.
The company was founded in 2021 by Yu Teng, a Tsinghua University alumnus, and initially focused on heterogeneous computing software before shifting its research focus toward quantum-inspired AI in 2025.
为什么这很重要
The company's approach aims to bridge the gap between current hardware limitations and the theoretical efficiency of quantum computing. By using quantum-inspired methods to compress models, Qingxing claims it can achieve high performance with significantly fewer parameters than traditional models. This is a notable development for enterprise AI, as it suggests a potential path toward deploying sophisticated, smaller-footprint models on existing, widely available hardware, rather than waiting for the maturation of general-purpose quantum computers. However, the practical utility of these models remains to be proven outside of specialized, narrow tasks.
The core value proposition of Qingxing's technology is the potential to reduce the computational overhead of AI models. By using quantum-inspired methods to optimize model parameters, the company seeks to lower the memory and processing requirements for deployment.
The reliance on conventional GPUs for deployment is a strategic choice that allows the company to bypass the current lack of accessible, general-purpose quantum hardware. This makes the technology immediately relevant to organizations that already possess standard GPU infrastructure.
The company's focus on 'token efficiency' and model structure redesign suggests an attempt to address the rising costs of AI inference. If successful, this could provide a more cost-effective alternative for companies looking to deploy specialized AI agents.
The lack of transparency regarding the third-party testing methodology and the specific data used for the RiverONE performance claims means that the industry cannot yet independently verify the superiority of this approach over traditional model compression techniques.
互动机制：它实际上是如何运作的
以交互方式探索这一发展背后的基础技术。
crm_get_transaction(id='4092').Which component of an AI application is the machine-learning model itself?
接下来看什么
The primary uncertainty lies in the scalability and generalizability of Qingxing's technology. While the company reports that its RiverONE model achieved 95% of the performance of a larger comparison model on a specific quantum task, it has not demonstrated similar results across broader, more diverse AI workloads. Furthermore, the company has not disclosed revenue figures or the scale of its commercial partnerships with chipmakers like Biren Technology and MetaX. Future updates should monitor whether the company can move beyond specialized technical calibration tasks to deliver consistent performance in general-purpose enterprise applications.
The company has not disclosed revenue, contract values, or the volume of repeat orders from its chipmaker partners, leaving the actual commercial viability of its business model unconfirmed.
Future performance benchmarks will be critical. The current evidence is limited to a narrow, specialized task (quantum chart interpretation). It remains unknown if these quantum-inspired methods maintain their efficiency and accuracy when applied to general-purpose tasks like natural language understanding or complex image generation.
The company's ability to maintain its technical edge while scaling its operations in the competitive Chinese AI market will be a key indicator of its long-term success.
Investors like Xuhui Capital have stated intentions to connect the company with local computing resources and industrial partners, which may accelerate the transition from private deployment to broader commercial availability.