Mind Lab: A Chinese Neo Lab Focused on Continuous Learning for AI
OpenAI
Anthropic
Alibaba/Qwen
Chinese startup Mind Lab, founded by serial entrepreneur Andrew Chen, is pioneering continuous learning and post-training for large AI models. Its Macaron-V1 model, built on GLM-5.2 with 4B LoRA parameters, achieved 6 SOTA results in benchmarks and reached $10M ARR within two weeks of commercialization. The company raised $60M in total funding, led by Meituan.
Mind Lab, founded in October 2025 by Andrew Chen, focuses on continuous learning and post-training for large language models. Its Macaron-V1-Preview, built on GLM-5.1 with five 1B-parameter LoRA expert modules, surpassed base models like GLM-5.1, GPT 5.4, and Claude Opus 4.6 in benchmarks. In July 2026, the official Macaron-V1 was released in two versions: Venti (748B total parameters, 4B trainable via LoRA) based on GLM-5.2, and Tall (50B) based on Qwen3.6, both supporting 200K token context. The company's approach uses Mixture-of-LoRA (MoL) for dynamic task adaptation and continuous model updates from user interactions. Mind Lab has built a post-training infrastructure platform called MinT, managing millions of LoRA models. The company has raised $60M in total funding, including a $50M A round led by Meituan in early 2026, with investors like Ant Group, Sequoia China, and ZhenFund. Mind Lab's technology aligns with the experiential intelligence concept, similar to Richard Sutton's experience-driven AI, and has been validated by independent research from Thinking Machines Lab. The company plans to offer continuous learning as a service for B2B clients, especially smart hardware makers.
- Сокращения
- MoL = Mixture of LoRA
- LoRA = Low-Rank Adaptation
- SOTA = State-of-the-Art
- MoE = Mixture of Experts
- RL = Reinforcement Learning
- Infra = Infrastructure
- MTP = Multi-Token Prediction
- DSA = Dynamic Sparse Attention
Source: 36Kr —
original