Mind Lab Tests Continual Learning with Macaron-V1

Mind Lab has made continual learning—where AI continuously improves through experience—its core direction, and developed Macaron-V1. The company was founded in October 2025 and is staffed by people from xAI, DeepMind, DeepSeek, ByteDance Seed, MIT, and Tsinghua University.

The Macaron-V1-Preview, released in June, combined GLM-5.1 with 5 LoRA expert modules of about 1 billion parameters each. Within two weeks of launch, it reached $10 million in annual recurring revenue (ARR), and on benchmarks presented by the company, it outperformed the base model as well as GPT-5.4 and Claude Opus 4.6.

The core technology is MoL (Mixture of LoRA adapters), which combines multiple LoRA adapters.

• It dynamically selects the most suitable expert module depending on the task type.
• It continuously updates by distilling data accumulated during use into a dedicated LoRA adapter.
• This improves per-user performance without retraining the entire model.

The official version, released on July 21, consists of two models. Venti is a 748 billion parameter model based on GLM-5.2, of which 744 billion parameters are the frozen base model and 4 billion are LoRA adapters trained by Mind Lab. The 4 adapters each handle conversation, agent, coding, and UI generation functions respectively.

Tall is a lightweight model for local deployment, based on Qwen 3.6 with 50 billion parameters. Both models natively support a 2 million token context window, and on benchmarks released by the company, Macaron-V1 achieved top-tier performance in 6 out of 12 tests.

Mind Lab extended the capabilities of GLM-5.2 by adding only 4 billion LoRA adapter parameters without significantly changing the overall parameters. This result shows that post-training and continual learning could become a major path to boosting next-generation model performance.