Norizon AI
Category: Foundation Models / LLMs
Beijing research-driven AI startup building models that keep learning after deployment, using a SHINE hypernetwork for real-time recursive self-improvement (RSI). Norizon AI was founded in 2026. The company is led by Muhan Zhang (Founder & Chief Scientist). Based in Beijing, China. Team size: 11-50. Total funding raised: Several hundred million RMB disclosed across angel and angel+ rounds; exact amount undisclosed.. Latest round: Angel. Key investors include Wuyuan Capital (五源资本), HongShan / Sequoia China (红杉中国), Legend Star (联想之星), Zhongguancun Science City Fund (中关村科学城基金), Gobi Partners (戈壁创投), Emerging Ventures (靖亚资本).
- Founded
- 2026
- Headquarters
- Beijing, China
- Team size
- 11-50
- Total funding
- Several hundred million RMB disclosed across angel and angel+ rounds; exact amount undisclosed.
Value proposition
Lets deployed LLMs absorb new knowledge and experience directly into their own parameters through a single forward-pass update, without repeated fine-tuning, long-context stuffing or external retrieval, while mitigating catastrophic forgetting.
Products and solutions
SHINE (Scalable Hyper In-context NEtwork), a hypernetwork that maps context to LoRA adapters for LLMs in one pass (ICML-26). The company is exploring enterprise model adaptation, agent experience learning and personalized models, with local and cloud delivery. No commercial product launched yet.
Unique value
Parameter-level continual learning at inference-level update cost, using a pre-trained hypernetwork to generate weight updates in one forward pass.
Target customer
Enterprises needing private-deployment models that absorb new business knowledge locally; AI agent developers; personalized-model and cloud-service use cases (all still exploratory).
Industries served
Enterprise AI, AI agents, personalized AI
Technology advantage
Pre-trained SHINE hypernetwork generating LoRA parameter updates in a single forward pass. Company claims better new-knowledge learning and general-task retention than fine-tuning baselines. Team is working on scaling the hypernetwork and stability across millions of tokens of continuous updates. Deep academic pedigree from the founder's PKU lab (MuLab).
How they differentiate
Targets the model-parameter layer of RSI, writing new knowledge into weights via a pre-trained hypernetwork. Sakana's Doc-to-LoRA uses a similar hypernetwork-to-LoRA idea but is research-stage. Many other RSI peers automate R&D workflow, chip design or research methodology instead.
Main competitors
Sakana AI (Doc-to-LoRA / Text-to-LoRA hypernetwork research), 超进化智能 (Super Evolution Intelligence), Recursive Superintelligence
Major milestones
2026-06: company founded in Beijing's Haidian district. 2026-07: SHINE paper accepted at ICML-26. 2026-11-19: announced several-hundred-million-RMB angel and angel+ financing.
Growth metrics
Team of 20+ people; incorporated June 2026; angel and angel+ rounds announced 2026-11-19.
Market positioning
Early-stage, research-led Chinese RSI / continual-learning startup backed by top-tier VCs. Pre-commercial, with no commercialization validation data disclosed.
Geographic focus
China (Beijing), with global RSI competition
Patents and IP
Published research: SHINE (ICML-26) and related hypernetwork continual-learning papers (e.g. arXiv 2610.15677). No patents found.
About Muhan Zhang (Founder & Chief Scientist)
Ex-Facebook AI (Meta AI) Research Scientist (2019-2021), worked on billion-user-scale graph ML systems; PhD Computer Science, Washington University in St. Louis (advisor Yixin Chen); BEng Shanghai Jiao Tong University (IEEE pilot class). Tenure-track Assistant Professor and PhD advisor at Peking University Institute for AI since 2021; area chair at NeurIPS/ICML/ICLR; Google Scholar citations over 14,000.