Mindful Tech's Mind Lab releases open-source 749B AI Agent Macaron-V1-Preview

On June 8, per BlockBeats' monitoring, Mindverse’s Mind Lab has open-sourced the personal AI model Macaron-V1-Preview, which has 749 billion parameters. Built on the GLM5.1 pre-trained base, it offers a 202k context window and is released under the MIT license. The development team, led by Chen Kaijie and Andrew (co-authors of the ReAct and FireAct papers), includes core researchers from DeepSeek, ByteDance, and xAI. Currently, the model has deep integration with Mindverse’s assistant app Macaron, which boasts 2 million users, creating a closed “Product-Model” iterative feedback loop. Earlier this June, Mindverse closed a nearly $50 million Series A funding round led by Meituan. Macaron-V1-Preview underwent reinforcement learning training based on GLM5.1. It adopts a Mixture-of-LoRA (MoL) architecture, where task-specific “LoRA skill packs”—covering emotion understanding, tool usage, and coding capabilities—are mounted on a shared base model. These skill packs support sub-second dynamic switching and data isolation. Because only lightweight skill packs are trained (not the 744 billion base model parameters), computational costs are significantly reduced. The team completed post-training for this 750 billion parameter model with fewer than 300 GPUs, at a computational cost less than 1% of that required for a model of similar size. The model supports Google’s A2UI protocol, which converts text responses into actionable web interactive interfaces in under 5 seconds. In evaluations across LivingBench, VitaBench, and smart assistant benchmark PinchBench, the model ranks first, with a PinchBench score of 92.5 points. The team has now initiated partnerships with leading mobile and wearable hardware manufacturers.