Meituan led the Series A funding round, bringing Mindverse's total funding to nearly $50 million. The company will open-source its first 750B reinforcement learning agent model based on GLM 5.1.

According to Beating, Mindverse, an AI agent developer, announced the completion of its Series A financing round, led by Meituan, with participation from Yuanhe Capital, Shokz, Variable Capital, and existing shareholders. The total financing reached nearly $50 million, with Gaohu Capital serving as the exclusive financial advisor. Mindverse was co-founded by Chen Kaijie and Andrew (who co-authored the ReAct/FireAct papers with Yao Shunyu). Its core R&D team comes from DeepSeek, ByteDance Seed, and xAI, and the company is dedicated to building native agent models through post-training and continuous learning technologies. Unlike conventional approaches that rely on external prompts, Mindverse firmly believes that the core capabilities of an agent come from the model's "post-training." The company uses LoRA as the technological foundation for continuous learning and parameterized memory, and by the end of 2025, it achieved lossless LoRA reinforcement learning for a trillion-parameter MoE model, reducing post-training costs 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 dock, supporting dynamic and rapid activation (in seconds) and clean privacy isolation. Mindverse announced that it will open-source its latest trained 750B parameter agent model, which includes a 744B pre-trained GLM 5.1 dock and 6B LoRAs. It is the world's first result to complete reinforcement learning training on GLM 5.1 and has achieved state-of-the-art (SOTA) results in four major leaderboards: Living Bench, Vita Bench, A2UI, and PinchBench. In terms of products, Mindverse adopts a two-way flywheel for collaborative iteration between model training and the C-end life agent application Macaron (2 million users, 100,000 daily active users); commercially, it has partnered deeply with Huawei Cloud, Microsoft Cloud, and others to open up low-cost, high-concurrency LoRA training and inference infrastructure to enterprises.