MiniMax introduced M3 with a one-million-token context window, five times the size advertised for M2.7. It also supports multimodal input, including images and video.
The company's SWE-Bench Pro results placed it above GPT-5.5 and Gemini 3.1 Pro and near Opus 4.7. Those launch comparisons need to be considered alongside independent tests and the requirements of an actual project.
My concerns about earlier MiniMax releases included exposed reasoning text and inconsistent adherence to instructions. I would test those behaviors in M3 rather than assuming that a larger context window resolves them.
The original Chinese launch rates were RMB 4.2 per million input tokens and RMB 16.8 per million output tokens below 500K context, with doubled rates for 500K–1M contexts. They are historical quoted rates, not a guarantee of current international pricing.
Source: MiniMax M3 announcement.
Adapted from the original Chinese article, published on June 1, 2026.
Comments NOTHING