Step on the key step of AGI mentioned by Liang Wenfeng! This model’s ARR exceeded 10 million US dollars in two weeks

Mind Lab Author | Cheng Qian Editor | Mo Ying Mind Lab reported on July 23 that on July 21, the Mind Lab team released its first official version of the MoE model Macaron-V1, including Macaron-V1-Venti and Macaron-V1-Tall, specially built for Agent tasks. Macaron-V1-Venti: It is the flagship model of 748B parameters, consisting of a basic model of 744B parameters and four 1B parameter LoRA expert modules. This is also the first model to be trained twice on GLM-5.2. Macaron-V1-Tall: Parameter size 50B, consisting of a basic model with 35B parameters and four LoRA expert modules with 3.7B parameters. The model is trained on Qwen 3.6 and is suitable for local deployment. By introducing the LoRA architecture, the overall performance of the Macaron-V1 series models is comparable to that of international head models in chat, programming, agent, and GenUI tasks, especially in long-range autonomous agent tasks. At the same time, Mind Lab launched commercialization in July this year, and in just two weeks, annualized recurring revenue (ARR) exceeded 10 million US dollars. All variants of the Macaron-V1 model are released as open weights in Hugging Face’s Macaron-V1 collection. 1. Comprehensive performance is benchmarked against top international models, and long-range autonomous agent capabilities have become the core watershed. AI applications have shifted from static question and answer to long-term autonomous agent applications, which also means that they no longer only compete for the capabilities of a single model, but instead cover system capabilities such as long-term memory, tool invocation, continuous self-optimization, multi-task collaboration, system architecture, and continuous learning capabilities. This is something the next generation model is missing too. According to an investor meeting recording circulated last night, DeepSeek founder and CEO Liang Wenfeng mentioned that the next generation model must have the ability to continue learning. He likened the route AGI will take to climbing a staircase. Last year