Heguang Quantum, a company based in Hangzhou, has completed a seed funding round and announced the creation of the first deterministic code for photonic quantum error correction. The company aims to sell the HQ10 quantum accelerator by 2027 for hybrid classical-quantum computing.
During the closing of the seed round, Heguang Quantum presented its development of a deterministic code for photonic quantum error correction, which solved the long-standing problem of reliable generation of GKP error correction resources via an attachable nonlinear hardware module.
The founder, Dr. Yu Shan, a PhD student of Academician Guo Guangkan who participated in the development of the world's first general-purpose large-scale photonic quantum computer and received funding from the EU's Marie Curie Fellowship program, built the company's approach around solving two main bottlenecks in quantum computing: scalability through distributed architecture and reliability through quantum error correction.
Instead of focusing on the industry-dominant metric of raw qubit count, Heguang Quantum concentrates on modular, expandable systems for photonic quantum computing. These systems can connect multiple quantum computing nodes via optical fiber networks, utilizing light's inherent advantage for both computation and communication.
The distributed architecture solves a fundamental physical limitation: as individual quantum computers grow, the complexity of control, system stability, and deployment costs increase exponentially. Heguang's approach connects multiple quantum nodes using photonic interconnects, allowing for larger effective computation without the need for a single monolithic device. The company has begun testing distributed quantum computing tasks at kilometer scale.
Concurrently, the HQ10 quantum acceleration processor is designed to integrate with classical computing resources, including GPUs, CPUs, and TPUs. It is positioned as a quantum accelerator module for specific application scenarios, rather than a complete replacement for existing classical computing infrastructure. GKP codes previously faced low success rates, probabilistic preparation, and scaling difficulties. The Heguang module provides deterministic GKP generation, moving error correction from laboratory verification to an applicable capability.
Product lines include the HQ10 quantum accelerator (sales in 2027), the full HA100 system, and the HC planning platform. The HQ10 is released as an accelerator module until full fault tolerance is achieved. Thus, Heguang demonstrates the industry's transition from counting qubits to systems engineering.
Observing the Universe means observing the past, since light, although it is the fastest known phenomenon, does not propagate instantaneously. The speed of light is approximately 300 thousand kilometers per second, allowing it to travel about 9.46 trillion kilometers in a period of one year. This vast distance is the focus of astronomical studies.
The rest of this article is exclusively available to members of the Digital View Club. To gain unrestricted access, it is necessary to subscribe to the Prime plan, which offers several advantages, such as ad-free browsing, access to exclusive content and newsletters, in addition to free courses and discounts on Prime courses, along with participation in an active community and special events.
The Chinese Academy of Sciences (CAS ICT) has introduced the ZhiJing social intelligence system, which incorporates the Zing model and the SoMBench evaluation benchmark. This development aims to address the issue of artificial intelligence's inability to comprehend social context, emotions, and unspoken signals.
Unlike models such as GPT-5.5, which demonstrate high performance in professional writing, or DeepSeek in coding, AI often produces awkward responses in social situations. ZhiJing treats social cognition as an independent engineering problem.
The first component of the system is SoMBench, which serves as a standard for evaluating social intelligence. It decomposes the abstract concept of human understanding into three main domains, 17 sub-dimensions, and 71 detailed tasks. These tasks cover relationship modeling, emotion comprehension, societal norms, and belief reasoning. The benchmark contains 3481 meticulously verified examples, utilizing multi-type cross-validation, including multiple-choice, single-choice, judgment, and open-ended questions, as well as option perturbation and label detection methods for precise error identification.
Testing across 20 leading models showed that the strongest model achieved only 72.08% accuracy, with none approaching the 90% threshold, confirming that social intelligence remains an under-researched area. The most complex challenges involve simultaneously reasoning about implicit intentions, emotional shifts, social norms, relational dynamics, and unstated words.
The Zing methodology includes three key innovations. The first is the FLARE data cycle: when weaknesses are identified, FLARE generates new samples targeted at closing this capability gap. The second is two-stage training, which first establishes a general foundation through distillation from multiple teachers and Mixed-Reward GRPO, followed by specialized reinforcement for emotions, intentions, and social norms. The third innovation, On-Policy Distillation (OPD), resolves the classic risk of catastrophic forgetting in reinforcement learning. OPD introduces constraints on teacher token distributions in online trajectories, allowing the training model to explore more efficient paths without deviating from established reasoning patterns. Thus, the framework operates holistically: FLARE determines what needs to be learned, two-stage training determines when to do it, and OPD determines who to learn from and what not to forget.
The Zing-27B model demonstrated a composite score across five benchmarks at 79.80%, surpassing GPT-5.5, which scored 78.44%. Zing-27B shows advantages in tracking recursive beliefs (HiToM) (4.08 percentage points higher) and EmoBench (5.62 percentage points higher). It is the first model under 100 billion parameters to outperform GPT-5.5 in the HiToM task, thereby asserting social intelligence as a distinct engineering discipline.
