Macaron-V1发布：两周ARR破千万美元，LoRA+MoE能否撬动AGI？
In the rapidly evolving landscape of artificial intelligence, Mind Lab has made a significant breakthrough with the release of its first official Mixture-of-Experts (MoE) model, Macaron-V1. Launched on July 21, 2024, this innovative series is specifically designed for Agent tasks and achieved remarkable commercial success, reaching an Annual Recurring Revenue (ARR) of over $10 million within just two weeks of its commercialization phase.
Architectural Innovation: LoRA + MoE Hybrid Design
The Macaron-V1 series introduces a novel approach to large language models through its unique architectural design:
Model Variants
-
Macaron-V1-Venti: The flagship version featuring 748 billion parameters, consisting of a 744B base model enhanced by four 1B LoRA expert modules. Built upon the GLM-5.2 foundation, this version represents the pinnacle of performance capabilities.
-
Macaron-V1-Tall: A more accessible 50-billion parameter version suitable for local deployment. Comprising a 35B base model with four 3.7B LoRA expert modules, it's trained on Qwen 3.6, offering practical solutions for organizations requiring on-premise AI capabilities.
Technical Breakthroughs
The core innovation lies in the Mixture-of-LoRA architecture, which offers several advantages:
- Incremental Continuous Learning: By freezing the base model and utilizing independent LoRA adapters, the system enables continuous learning without catastrophic forgetting
- Significant Cost Reduction: Training and inference costs are reduced to approximately one-tenth of traditional full-parameter fine-tuning methods
- Competitive Performance: In areas such as chat, programming, Agent tasks, and GenUI, Macaron-V1-Venti demonstrates comprehensive performance comparable to leading international models like GPT-5.5 and Claude Opus 4.8, particularly excelling in long-horizon autonomous agent tasks
Infrastructure and Tooling
To support this new paradigm, Mind Lab introduced the MindLab Toolkit (MinT), a sophisticated platform capable of managing millions of LoRA strategy resources. This infrastructure enables efficient online inference services while sharing trillion-parameter base models, democratizing access to cutting-edge AI capabilities.
Commercial Success and Strategic Partnerships
The market response to Macaron-V1 has been exceptionally positive:
- Strong Funding Backing: In June 2024, the company secured nearly $50 million in financing from prominent investors including Meituan, Yuanhe Puhua, Shokz, and Variable Capital
- Rapid Revenue Growth: Within two weeks of launching commercial operations in July 2024, the ARR exceeded $10 million
- Strategic Collaborations: The company has partnered with industry giants ByteDance and NVIDIA to build a LoRA training foundation, while also collaborating with AI hardware companies like Shokz
Industry Implications
This development aligns with insights from DeepSeek founder Liang Wenfeng, who emphasized that next-generation models must possess continuous learning capabilities—a crucial step toward achieving Artificial General Intelligence (AGI).
The "strong shared base + massive evolvable lightweight adapters" architecture exemplified by Macaron-V1 represents a practical engineering implementation of this theoretical framework. By enabling models to continuously learn and adapt without retraining entire systems from scratch, this approach could fundamentally transform how organizations develop and deploy AI solutions.
As the AI industry continues to evolve, innovations like Macaron-V1 demonstrate the growing importance of efficient, scalable, and adaptable model architectures in the quest for more capable and accessible artificial intelligence systems.