MiniMax Raises About $2B and Releases M3 Model Usable Inside FiftyOne

Shanghai-based AI lab MiniMax, listed in Hong Kong since January, has announced a private placement and convertible-bond raise of roughly HK$16 billion, its first major fundraising since the IPO, with most funds directed to model research and infrastructure. Alongside this, the company released MiniMax-M3, a natively multimodal model with about 428 billion total parameters and a 1 million-token context window. A new FiftyOne plugin lets users apply the model to images and videos for detection, keypoints, captioning, visual question answering and other tasks without fine-tuning or hosting a model server. The plugin writes outputs as native FiftyOne label objects and records provenance such as prompts and sampling settings. The model uses a sparse attention mechanism that its developers report substantially cuts compute and speeds processing at long context lengths.