Not just because of the benchmark claims, although those are loud: MiniMax says M3 hits frontier-level coding and agent performance, supports up to 1M tokens of context, and is natively multimodal across text, image, video, and computer use.
The part I care about most:
We are moving from “AI that answers prompts” to “AI that can stay with a project.”
Long context matters because real work is messy. A repo, a product brief, old decisions, tool logs, screenshots, failed attempts, user corrections, and new constraints all need to live in the same working memory.
Multimodality matters because real work is not just text. It is docs, code, screenshots, dashboards, PDFs, browser state, apps, and weird edge cases.
Agentic coding matters because the valuable use case is not “write me a function.” It is: understand the system, make a plan, change the code, test it, notice what broke, and keep going.
MiniMax also says M3 is open-weight, with the technical report and weights expected soon. If that lands well, it could put more frontier-style capability into the hands of builders who do not want every workflow locked inside one closed platform.
The next AI advantage will not just be who has the best chatbot, but who builds the best operating system around persistent, multimodal, tool-using intelligence.