Amid the slowing of the general large model红利, AI applications are at a critical turning point, transitioning from "scalability" to "personalization." On January 27, 2026, Mindverse, a company dedicated to the development of personal intelligence models, announced a Pre-A round financing exceeding $20 million.
This round of financing was led by Ant Group, with joint investments from institutions such as Sequoia China, marking the official transition of "personal intelligent models" from conceptual research to large-scale implementation.
Technical Breakthroughs: Trillion-Parameter LoRA and Reinforcement Learning
Mindverse believes that true "individualized" understanding of users should not rely on complex prompts or external knowledge bases, but rather on models that possess continuous growth, individual personalities, and long-term memory.
LoRA Core Architecture: Mindverse uses LoRA as a parameter layer for sustainable training, carrying users' long-term memory and personality preferences, enabling the model to continue evolving based on feedback after pre-training.
Engineering Leadership: The team has already achieved a large-scale LoRA reinforcement learning system and built the trillion-parameter training platform Mind Lab Toolkit. Currently, only a few teams globally have this capability, including Mindverse and Thinking Machines Lab, founded by the former CTO of OpenAI.
Product Practice: 10x Evolution of Macaron AI
As an experimental product of Mindverse, Macaron AI has demonstrated strong implementation capabilities in personal life scenarios:
Extreme Speed: The application generation time has been reduced from 20 minutes initially to within 10 seconds; through its self-developed container solution, the startup time has been optimized from 40 seconds to about 2 seconds.
High Reliability: System operation success rate has increased from 90% to 99.9%.
Ecosystem Scale: The number of small applications created by users within the platform has exceeded 300,000.
Future Vision: From 28 LoRAs to "One Model per Person"
Currently, Mindverse is running 28 different LoRA models based on user types. The company's next goal is to provide a customized model for tens of thousands of users by Q2 2026, truly achieving the vision of "personal intelligent agents."
