en.Wedoany.com Reported - On July 10, MiniMax completed a new funding round totaling approximately 16 billion HKD. The transaction attracted over a hundred institutions, achieving a 7x oversubscription, with the issuance scale expanding from an initial ~$1.8 billion to over $2 billion. The capital injection will primarily serve AI infrastructure, model research and development, agent engineering, and global product operations.
This funding round consists of two parts: a new share placement and zero-coupon convertible bonds. The placement of 35.6 million new Class A shares is expected to raise approximately 9.54 billion HKD, while the zero-coupon convertible bonds amount to 6.5 billion HKD. MiniMax plans to allocate about 80% of the proceeds to AI infrastructure and model R&D. The use of funds is not for general business expansion, but to continue increasing investment in computing clusters, model training, inference services, and agent product engineering.
Large model R&D requires continuous handling of two types of computing demands: training and inference. The training phase involves massive data processing, distributed computing, model parameter updates, and multiple rounds of experimentation. The inference phase requires the model to maintain low latency, high concurrency, and stable output during real user calls. As context windows expand and models enter multimodal and agent tasks, the text, images, videos, code, tool return results, and historical states that need to be processed in a single task continue to increase, requiring synchronized expansion of computing power, storage, networking, and engineering scheduling.
MiniMax's latest flagship model, M3, has concentrated its technical direction on coding, agents, ultra-long context, and native multimodality. The model adopts the MiniMax Sparse Attention architecture, supporting up to 1 million Token context and capable of processing image and video inputs and operating desktop environments. Under a 1 million Token context, M3's per-Token computation is reduced to approximately one-twentieth of the previous generation model, with prefill speed increased by over 9 times and decoding speed by over 15 times.
Agent engineering is another main line of this capital investment. For a model to accomplish real work, it cannot stop at generating a piece of text or code; it also needs to understand goals, break down tasks, call search and office tools, read files, execute programs, check results, and correct subsequent actions based on intermediate feedback. Agent engineering requires solving tool interfaces, task states, permission control, long-term memory, error recovery, and execution evaluation, enabling the model to work continuously in longer task chains.
M3 has extended its coding capabilities to program bug fixing, front-end and back-end development, performance optimization, and terminal execution. Through an interactive user simulation framework, the model is trained to handle requirement clarification, solution discussion, feedback modification, and complex project iteration. The model no longer only completes single-turn instructions but attempts to collaboratively advance software tasks with developers in continuous sessions.
MiniMax has now formed a system of language, video, speech, image, and music models, and has launched code agents, video generation, audio generation, AI character applications, and an enterprise development platform. With the injection of 16 billion HKD, AI infrastructure expansion, M3 and subsequent model training, agent engineering, long-context computing optimization, and multimodal product deployment will become specific investment areas.
