Skip to main content

Mind Lab

Category: AI Agents

Mind Lab is the frontier AI research arm of Macaron AI (Mindverse/心洲科技), focused on Experiential Intelligence — building AI agents that learn and improve through real-world reinforcement learning using LoRA-based post-training. Mind Lab was founded in 2025. The company is led by Kaijie Chen (陈锴杰). Based in Beijing, China. Team size: 11-50. Total funding raised: $50M. Latest round: Series A. Key investors include Ant Group, Meituan, Hongshan (Sequoia China), ZhenFund, Linear Capital, Gaorong Capital, Yuan One Capital, Yuanhe Puhua, Shaoyin Technology, Variable Capital.

Founded
2025
Headquarters
Beijing, China
Team size
11-50
Total funding
$50M

Value proposition

Mind Lab enables AI agents to learn continually from real-world experience through LoRA-based reinforcement learning at 10% of typical GPU cost, powering Macaron AI's Personal Agent that creates personalized mini-apps for everyday life.

Products and solutions

Macaron-V1 (748B Venti & 50B Tall agent models, first post-trained on GLM-5.2), MinT (Mind Lab Toolkit - RL infrastructure for trillion-parameter LLMs), Macaron AI (Personal Agent app - creates personalized mini-apps for daily life), LongStraw (long-context RL post-training up to 2M tokens), Mixture-of-LoRA (MoL) architecture, UI4A (Generative UI solution), Macaron ChatBench & LivingBench (Personal Intelligence benchmarks)

Unique value

First team globally to run high-performance RL on trillion-parameter open-source models using LoRA at 10% GPU cost; Mixture-of-LoRA architecture enabling "one model per person" personalization; Research-product co-design loop between Mind Lab and Macaron AI app.

Target customer

Individual consumers seeking personalized AI life assistants (B2C via Macaron AI app); Enterprise/developer teams needing customized agent models via LoRA post-training (B2B via MinT platform)

Industries served

Personal AI Assistants, Consumer Lifestyle, Generative UI, AI Agent Infrastructure

How they differentiate

Unlike competitors focused on pre-training scale or productivity tools, Mind Lab/Macaron AI focuses on post-training via RL + LoRA for personalization. They pioneered trillion-parameter LoRA RL (only 3 teams globally: Mind Lab, Thinking Machines Lab, Fireworks). Their models are purpose-built for personal/lifestyle agent tasks rather than general-purpose benchmarks. They achieve SOTA on Personal Intelligence benchmarks (LivingBench, VitaBench, A2UI, PinchBench) while being open and self-hostable.

Main competitors

Thinking Machines Lab (Mira Murati's startup, same LoRA RL approach), Lovable (AI app builder), Cursor (AI coding), Anthropic's Claude (general agent capabilities), ByteDance (VolcEngine/VERL)

Key partnerships

NVIDIA (techniques merged into NeMo Megatron-Bridge), ByteDance (techniques merged into VolcEngine RL / VERL), Zhipu AI (GLM-5.2 as base model for Macaron-V1-Venti), Qwen (Qwen 3.6 as base model for Macaron-V1-Tall)

Notable customers

200万+ users (2M+ users) on Macaron AI app, 100K+ DAU, 300K+ user-created mini-apps on platform

Major milestones

Dec 2025: Mind Lab officially launched, First trillion-parameter LoRA RL achieved at 10% GPU cost, Techniques merged into NVIDIA Megatron & ByteDance VERL, Aug 2025: Macaron AI product launched, Jan 2026: $20M+ Pre-Series A led by Ant Group, Jun 2026: $30M Series A led by Meituan ($50M total), Jul 2026: Macaron-V1 released (748B/50B), first model post-trained on GLM-5.2

Growth metrics

200万+ users; 100K+ DAU; 300K+ user-created apps; App generation time reduced from 20 min to 10 seconds; 99.9% success rate

Market positioning

Pioneer in "Personal Intelligence" — differentiated from productivity AI by focusing on lifestyle/personal use cases. Positioned as a research-first company (Mind Lab) that feeds into a consumer product (Macaron AI), similar to how DeepMind feeds into Google or Anthropic's research feeds Claude. Backed by top-tier Chinese investors (Ant Group, Meituan, Sequoia China).

Geographic focus

China (primary); Singapore (HQ for Macaron AI); Global (open-source model releases on HuggingFace)

Patents and IP

Mixture-of-LoRA (MoL) architecture; LongStraw long-context RL training infrastructure; MinT RL platform; Macaron ChatBench & LivingBench evaluation frameworks; UI4A Generative UI system

About Kaijie Chen (陈锴杰)

Ex-MidReal (AI storytelling platform, 2M+ users) Founder; Ex-Yunzhongzi Technology (AI gamified social) Founder; Duke University Mechanical Engineering (HCI), GPA 3.99; Gold Medal Yau Science Awards Physics 2013; Intern at Zhihu under CTO Li Shenshen

Latest news about Mind Lab

More AI Agents companies

Official website:

Mind Lab - AI Startup Profile | AI Market Watch