Mind Lab under Mindverse Open-Sources 749B Intelligent Agent Macaron-V1-Preview

BlockBeats News, June 8: According to monitoring by Beating, Mind Lab, a subsidiary of Mindverse, has open-sourced its personal intelligent agent model Macaron-V1-Preview, which has 749 billion parameters. The model is based on the GLM5.1 pre-training framework and provides a 202K context window, open-sourced under the MIT license. The development team was co-founded by Kai-Jie Chen and Andrew, a collaborator on the ReAct and FireAct papers, with core R&D members from companies such as DeepSeek, ByteDance, and xAI. Currently, the model has been deeply integrated with Macaron, an assistant application under Mindverse with 2 million users, forming a "product-model" iterative feedback loop. Earlier this June, Mindverse completed a $50 million Series A funding round led by Meituan. Macaron-V1-Preview underwent reinforcement learning fine-tuning on GLM5.1. The model adopts a Mixture-of-LoRA (MoL) architecture, encapsulating task-specific skills such as emotion, tools, and code into independent "LoRA skill packs" mounted on a shared base. These skill packs support second-level dynamic switching and data isolation. Since the training process does not require adjusting the 744B base parameters and only involves training lightweight skill packs, computational costs have been significantly reduced. The team completed the fine-tuning of the 750B parameter model using fewer than 300 GPUs, with computational costs less than 1% of models of the same size. The model supports Google's A2UI protocol, enabling the transformation of text responses into actionable web interactive interfaces within 5 seconds. In evaluations such as LivingBench, VitaBench, and the intelligent assistant benchmark PinchBench, the model ranked first across the board, achieving a score of 92.5 in the PinchBench evaluation. The team has already begun discussions for collaboration with leading smartphone and wearable hardware manufacturers. [Original Link]