RSI Company Norizon AI Completes RMB 100 Million Seed Round | 5Y News
Generate parameter updates in a single forward pass using a hypernetwork, letting the model learn as it goes like a "hippocampus," with update costs close to those of inference.


Beijing Wuya Zhiyuan Technology Co., Ltd. (Norizon AI), a global leader in continuous parameter learning for large models, has announced the completion of a nine-figure RMB angel and angel-plus funding round. The round drew participation from several prominent investors including 5Y Capital, HSG, Legend Star, the Zhongguancun Science City Fund, and Gobi Partners. The proceeds will be used primarily to advance the company's in-house RSI model technology and bring its products to market.
Letting AI keep evolving after deployment
Norizon AI focuses on Real-Time RSI (real-time self-evolution), with continuous parameter learning as its core technical approach. The goal is to let large models keep absorbing knowledge, internalizing experience, and acquiring new capabilities after they have been deployed.
As AI agents take on complex, long-horizon tasks, autonomous exploration can generate more data and execution trajectories — but gains in exploration efficiency don't automatically translate into model capability growth. Absorbing new experience quickly, lowering the cost of model training, and retaining existing abilities while learning new knowledge are the key hurdles continuous learning must clear. Norizon AI's R&D centers on fast updates, inference-level update costs, and mitigating catastrophic forgetting — building the underlying learning capability that makes real-time self-evolution possible.
The technical vision takes human continuous learning and memory consolidation as a functional reference, exploring a learning mechanism akin to a "hippocampus" for models — converting information and experience worth keeping from interactions into parameter updates that can be reused long-term. By folding learning into everyday operation, the company hopes to enable continuous accumulation of model capability and reduce reliance on periodic, concentrated training and version iteration.
On the technical side, the company uses a pretrained hypernetwork to directly generate model parameter updates. A hypernetwork is a network capable of generating some or all of the parameters of another neural network: it takes as input task and data characteristics or model state, and outputs the weight updates the target network needs. Through the hypernetwork, parameter updates for the base model are generated directly, so new knowledge becomes part of the model itself — deposited into its own parameters and usable in subsequent tasks.
The team has already validated this approach on tasks requiring internalization of new materials: a single forward computation generates a lightweight parameter update, and the adapted model can answer related questions without directly reading the source text. This method avoids repeated gradient optimization for each new document, bringing the cost of a single parameter update close to inference level and laying a paradigm foundation for high-frequency, low-cost model updates.
On mitigating catastrophic forgetting, the hypernetwork outperformed fine-tuning controls on both new-knowledge learning and general-task performance. Building on this, the company is further studying scaling up the hypernetwork and the length of a single internalization pass, the stability of continuous updates across millions of tokens, actively discriminative and safely rollback-able internalization mechanisms, and more — pushing the model to absorb vast amounts of new experience while preserving base capabilities, and verifying that learning gains persist on longer-horizon tasks.
Norizon AI has assembled a core team spanning research, engineering, and product. Its chief scientist is Dr. Muham Zhang, currently an assistant professor at Peking University's Institute for Artificial Intelligence. He holds a bachelor's degree from the IEEE pilot class at Shanghai Jiao Tong University and a PhD from Washington University in St. Louis, and previously served as a research scientist at Facebook AI (now Meta AI), where he led the development of large-scale graph machine learning systems serving a billion users. Norizon AI has gathered a group of young researchers, including talent from Peking University, Tsinghua University, the University of Hong Kong, Fudan University, and MIT, as well as gold medalists in olympiad competitions and the ICPC, and researchers from leading tech companies working on post-training and agents. The team numbers more than 20 people.
As a frontier explorer of continuous parameter learning and real-time self-evolving models worldwide, Norizon AI is exploring applications including AI agent experience learning, enterprise model adaptation, and personalized models. In enterprise settings, the company aims to use private deployment so models can absorb new business knowledge locally, lowering the barrier to updates and reducing the leakage of sensitive data; in agent settings, it is exploring how effective experience from conversations, work records, and task trajectories can be applied to subsequent tasks, so the model better adapts to users' working styles and actual needs.
On delivery, the team is exploring both on-premise deployment and cloud services. In the cloud, a shared general-purpose base with personalized parameters maintained for each user can lower the cost of serving dedicated models. The team will progressively validate the product form and commercial value of continuous learning around these real-world needs.
The company's long-term goal is to build an intelligent system that can autonomously accumulate experience, distill transferable patterns, and continuously acquire new capabilities — making real-time self-evolution a foundational capability of models. Norizon AI wants learning to run through the entire lifecycle of a model, extending improvement from periodic version releases to everyday use, so that models keep adapting to new tasks and environments through long-term practice in the real world.


