【AI前沿】#2China's Multimodal Large Model Reaches a Milestone: MiniMax M3 Is Officially Open-Sourced and Response Speed Doubles

AI NEWSLatest AI NewsArticleChina’s Multimodal Large Model Reaches a Milestone: MiniMax M3 Is Officially Open-Sourced and Response Speed DoublesPublished in Latest AI NewsTime :Jun 16, 2026Read :3minuteMiniMax, a well-known technology company in the field of artificial intelligence in China, officially announced the open source of its native multimodal flagship model MiniMax M3 today. Previously, the official team had already released the weights of the model to the industry in advance and simultaneously published a highly technical paper on the sparse attention mechanism, which has attracted widespread attention from the open-source community.Outstanding Performance Leads Open SourceAs the flagship model developed by MiniMax, M3 has a total parameter count of up to 428B, with 23B active parameters. Notably, this model is the first open-source large model in the industry that directly performs multimodal mixed training from the initial stage of training.Thanks to its outstanding technical architecture, M3 quickly secured the top position among open-source models on the authoritative global comprehensive intelligence index ranking within just two weeks of its release. In response to the initial surge in traffic causing experience delays, the official team has optimized and improved the model’s output speed from an initial 30 TPS to 80 TPS, and further improvements are planned for the future.Top-Notch Coding and Code Delivery CapabilitiesIn various high-difficulty coding and intelligent agent capability evaluations, MiniMax M3 has demonstrated industry-leading technical standards. It not only has the ability to autonomously break down complex tasks but can also skillfully use various external tools for multi-step reasoning.Compared to traditional models, M3’s code writing capabilities have made a qualitative leap. The official team stated that the code it generates is intended to meet deliverable standards directly, rather than being a half-finished product that barely runs but requires extensive manual modifications, which will greatly enhance developers’ productivity.