Mind Lab puts continual learning to the test with Macaron-V1
Mind Lab, a startup founded by serial entrepreneur Chen Kaijie, has released Macaron-V1, a continual learning model that achieved state-of-the-art results in several benchmarks. The company's approach involves post-training large language models using a mixture of LoRA adapters, allowing for dynamic adaptation to different tasks and continuous improvement through user interaction data. Mind Lab's technology, which enables reinforcement learning at the trillion-parameter scale with significantly reduced GPU resources, has attracted USD 60 million in funding, including a Series A round led by Meituan's investment arm. This innovation positions Mind Lab as a key player in advancing AI's ability to learn and evolve from experience, moving beyond static, labeled-data paradigms.
Mind Lab's advancements in continual learning and LoRA-based post-training represent a significant shift in Asia's AI landscape. By enabling models to adapt and improve continuously with user data, the company addresses a critical limitation of traditional LLMs, which struggle with domain-specific use cases once training is complete. This technical direction, championed by industry leaders like Richard Sutton and DeepSeek, is becoming a consensus in the AI community, signaling a future where AI models are not static but evolve in real-time. Mind Lab's success in making LoRA-RL work on trillion-parameter models, particularly on GLM-5 series architectures, highlights a unique technical moat within China, differentiating it from global counterparts like Thinking Machines Lab.
The rapid commercial validation, with USD 10 million in annual recurring revenue just two weeks after launch, underscores the market's demand for adaptable and efficient AI solutions. The company's strategic focus on both consumer applications like Macaron and enterprise offerings through its MinT platform positions it to capture diverse segments of the AI market. This dual approach not only generates valuable user interaction data for model training but also provides a scalable infrastructure for other businesses to leverage continual learning, accelerating the adoption of this advanced AI paradigm across Asia.
