Meituan Leads Series A Funding Round, Mindverse Raises Nearly $50 million in Total, to Open Source the First 750B Reinforcement Learning Agent Model Based on GLM 5.1

According to WatchTower Beating monitoring, AI intelligence R&D company Mindverse announced the completion of a Series A funding round led by Meituan, with Yuanhe Bohai, ShawYin, Variant Capital, and existing shareholders participating, raising nearly $50 million in total, with GaoHu Capital serving as the exclusive financial advisor. Mindverse, co-founded by Chen Kaijie and Andrew (who previously collaborated with Yao Shunyu on the ReAct/FireAct paper), has a core R&D team from DeepSeek, Byte Seed, xAI, etc., dedicated to building native agent models using post-training and continual learning technologies.

Unlike the conventional practice of relying on external prompts, Mindverse believes that the core capability of an agent comes from the model's "post-training." The company has adopted LoRA as the technological foundation for continuous learning and parameterized memory. By the end of 2025, they achieved lossless LoRA reinforcement learning of a trillion-parameter MoE model, reducing the cost of post-training to one-tenth of full-parameter training. Through the Mixture of LoRA architecture, thousands of independent "skill packs" (LoRAs) representing different preferences or business experiences can be mounted on a shared base to support dynamic rapid activation (in seconds) and clean privacy isolation.

Mindverse announced plans to open-source its latest 750B parameter Agent model, which includes a 744B pre-trained GLM 5.1 base and 6B LoRA. It is the world's first achievement of post-training reinforcement learning on GLM 5.1, achieving SOTA in the Living Bench, Vita Bench, A2UI, and PinchBench rankings. In terms of products, Mindverse adopts a dual-cycle approach of model training and the C-end lifestyle Agent application Macaron (2 million users, 100,000 daily active users) for synergistic iteration; commercially, it deepens cooperation with Huawei Cloud, Microsoft Cloud, etc., opening up low-cost, high-concurrency LoRA training and inference infrastructure to enterprises.