Mind Lab builds Mint Recursive to help companies train their own AI
Mind Lab has launched Mint Recursive, a post-training and inference platform designed to help companies build and customize AI models using their proprietary business data. The platform also unveiled Macaron-V1.1, a 752-billion-parameter model post-trained on Mint Recursive.
Intelligence analysis by Gemini 2.5 Flash
Amid a shift in investor focus back to AI infrastructure, Mind Lab introduced Mint Recursive, a platform enabling enterprises to fine-tune large language models with their specific business experience. This aims to make AI more practical and cost-effective for industry use, moving beyond generic applications.
Imagine you have a super smart robot that knows a lot, but you want it to be really good at *your* specific chores, like sorting your toys or helping with homework. Mind Lab built a special workshop called Mint Recursive. It helps companies teach their big, smart AI robots to become experts at *their* unique jobs, like understanding their customers or writing special computer code, without having to buy a whole new robot every time. It's like giving the robot special training wheels that help it learn faster and cheaper.
Analysis
The launch of Mint Recursive by Mind Lab signifies a crucial evolution in the artificial intelligence sector, particularly within China. As investor enthusiasm for general AI applications cools, there's a clear pivot towards the foundational infrastructure that supports the development and deployment of large models. Mind Lab's platform directly caters to this shift, providing tools for post-training and inference that are essential for companies looking to integrate AI deeply into their operations.
Mint Recursive
Mint Recursive is presented as a comprehensive platform covering data management, evaluation, post-training, deployment, and inference. Its design allows users to access training and deployment tools via APIs and a Python SDK, making it accessible for developers. A key feature is its ability to collect feedback from models in use, creating a continuous learning loop that allows AI to adapt and improve based on real-world business tasks. This closed-loop system is vital for developing AI that is truly tailored and effective for specific enterprise needs.
The platform supports various advanced training approaches, including supervised fine-tuning (SFT), reinforcement learning (RL), and direct preference optimization (DPO), and is compatible with a range of open-source models like GLM, Qwen, and DeepSeek. By offering serverless training and deployment, Mint Recursive eliminates the need for companies to invest in expensive GPUs or manage complex computing clusters, instead charging based on token usage. This significantly lowers the barrier to entry for businesses wanting to leverage advanced AI capabilities without substantial upfront hardware costs.
Macaron-V1.1
Mind Lab showcased the capabilities of Mint Recursive through its own model, Macaron-V1.1, a 752-billion-parameter model post-trained from GLM-5.3. This model serves as a tangible demonstration of the platform's efficiency, with its latest update taking less than two weeks. Macaron-V1.1 is structured with a 744-billion-parameter base model augmented by four two-billion-parameter expert modules, each specializing in areas like chat, agents, coding, and generation.
The performance of Macaron-V1.1, particularly its strong showing on the SWE-Marathon benchmark for extended coding tasks, highlights the effectiveness of Mind Lab's approach to specialized AI. The rapid iteration cycle, moving from V1-Preview to V1.1 in a short span, underscores the platform's ability to accelerate model development and refinement. This model not only validates Mint Recursive's technical prowess but also provides a powerful example of how specialized AI can be built and optimized for complex tasks.
LoRA
Low-rank adaptation (LoRA) is a cornerstone technology for Mint Recursive, enabling efficient and cost-effective post-training. LoRA allows models to learn from business feedback and improve at specific tasks by training only a small set of additional parameters, rather than the entire model. This method has been shown to match the performance of full-parameter fine-tuning when configured correctly, as highlighted by research from Thinking Machines Lab.
Mind Lab's long-standing focus on LoRA is evident in Mint Recursive's architecture, where each specialist module operates as a lightweight adapter attached to a shared base model. This modularity means adapters can be upgraded, launched, or taken offline independently without affecting others, offering unprecedented flexibility and control. The ability to merge an adapter into the base model to create a standalone version further enhances customization, allowing companies to target specific business needs while substantially reducing post-training, deployment, and update costs.
Key points
- Mind Lab launched Mint Recursive, a platform for post-training and inference of AI models for industry use.
- The platform enables companies to customize AI models using their proprietary business experience and data.
- Mind Lab also unveiled Macaron-V1.1, a 752-billion-parameter model post-trained on Mint Recursive, demonstrating its capabilities.
- Mint Recursive supports various training approaches and open-source models, offering serverless deployment and feedback collection for continuous learning.
- The platform leverages LoRA (low-rank adaptation) to reduce post-training costs and allow for flexible, independent specialist modules.
Mind Lab's Mint Recursive platform could significantly democratize advanced AI customization for businesses, allowing them to leverage proprietary data to build highly specialized models without prohibitive costs or complex infrastructure management. This could lead to a surge in innovative AI applications tailored to specific industry needs, boosting efficiency and competitiveness across various sectors.
Despite the promise, the success of Mint Recursive hinges on companies' ability to provide high-quality, relevant proprietary data, which can be a significant challenge. Furthermore, the complexity of integrating and managing these customized AI workflows, even with a simplified platform, might still pose adoption hurdles for smaller or less tech-savvy enterprises.
