MiniMax M3, an open-source, native multimodal flagship model.

MiniMax M3, the open-source native multimodal flagship model, boasts 428 bytes of total parameters and 23 bytes of activation parameters. Employing a self-developed MSA sparse attention architecture to reduce the computational cost of long contexts, it is the first open-source model to undergo multimodal hybrid training from Step 0. Since its release, it has achieved first place among open-source models on multiple leaderboards, including Artificial Analysis Comprehensive Intelligence Index, GDPval-AA, and Code Arena WebDev, and has received industry acclaim from the CEO of Vercel and the founder of Y Combinator. (Source: Xiyu Technology)