China's Neo Lab Focuses on Continuous Learning for Next-Gen AI Models
Chinese AI startup Neo Lab, also known as Mind Lab, is pioneering continuous learning techniques for large language models, aligning with prominent researchers like Richard Sutton, a Turing Award laureate and father of reinforcement learning. The company, founded in October 2025 by serial entrepreneur Chen Kaijie, comprises around 30 individuals with backgrounds from institutions like xAI, DeepMind, and MIT. Neo Lab's approach emphasizes experience-driven AI, moving beyond static, labeled datasets which are seen as reaching their limits. Their recent Macaron-V1-Preview model, built on GLM-5.1 with LoRA expert modules, demonstrated superior performance over established models like GPT-5.4 and Claude Opus 4.6. The core innovation lies in their Mixture-of-LoRA (MoL) method, enabling dynamic switching of specialized modules based on task requirements and continuous updating with user data. This strategy has shown rapid commercial success, achieving $10 million in Annual Recurring Revenue (ARR) within two weeks of its initial business rollout. The formal release of Macaron-V1 on July 21st includes a flagship 748B parameter model, Venti, and a lightweight 50B parameter model, Tall, both supporting a 2 million token context window. Neo Lab has secured over $60 million in funding, with a recent $50 million Series A led by Meituan. The company views continuous learning as a fundamental shift, enabling models to adapt and evolve from their own experiences, a concept they term 'Experiential Intelligence'.
Neo Lab's focus on continuous learning and Mixture-of-LoRA (MoL) represents a strategic pivot towards adaptive AI, addressing the limitations of static large language models. By leveraging LoRA (Low-Rank Adaptation) for efficient fine-tuning and MoL for dynamic specialization, the company aims to create models that evolve with user interaction and task demands. This approach aligns with a broader industry trend recognizing the need for AI systems to learn and adapt continuously, rather than relying solely on periodic, large-scale retraining. The commercial success achieved by Neo Lab suggests a market readiness for these more agile AI solutions. However, the long-term scalability and robustness of MoL architectures, particularly in managing potentially millions of LoRA modules and ensuring seamless collaboration between them, will be critical. Future developments will likely explore optimal LoRA collaboration strategies and the infrastructure required to support this dynamic learning paradigm, potentially influencing how AI systems are deployed and maintained in the coming decade.
AI-generated to prompt reflection — not editorial opinion, not advice, not a statement of fact. How this works.