MiniMax teases upcoming M3 model with new sparse attention mechanism and 15.6X long-context response speed boost

MiniMax released a technical report on its M2 family, outlining engineering choices behind its open-source language models across text, coding, and agent use cases. The company also teased its upcoming M3 models, saying a new sparse attention approach can deliver up to 15.6x faster decoding at million-token context lengths.

The announcement highlights growing competition among Chinese AI labs to push frontier capabilities while keeping models accessible to developers and enterprises. If MiniMax’s M3 claims hold up, long-context agents could become cheaper and more practical for production workflows that require large document sets, memory, or tool use.