All newsReleases

MiniMax built its name on a million tokens — then spent 220 days selling one fifth of that

Published: 8/24/2026 · Source: MiniMax — dokumentacja platformy i karty modeli na Hugging Face

MiniMax entered the market with the longest context window anyone was offering. MiniMax-Text-01, published on 15 January 2025, was trained on a one-million-token window and documented up to four million tokens at inference. MiniMax-M1, the reasoning model that followed in June 2025, kept the million-token input. That number was the company's calling card. Then it went away. MiniMax M2, released on 23 October 2025, came with a context window of 196,608 tokens — roughly one fifth of what the same company had shipped nine months earlier. M2.1 (22 December 2025) and M2.5 (12 February 2026) kept exactly the same figure. M2.7 (18 March 2026) lifted it to 204,800, an increase of four percent. The million-token window returned only with M3 on 31 May 2026, which documents 1,048,576 positions. That is 220 days — from M2 to M3 — during which the newest model MiniMax sold had a context window five times smaller than the model it had published in January 2025. And the older model was still on sale: M1, with its million-token input, remained available at independent providers throughout, which meant a buyer picking "the latest MiniMax" for a long-document task was picking the worse tool for it. The cut was not an accident. The M2 line is a different animal: a sparse mixture of experts with 230 billion total parameters but only about 10 billion active per token, built for coding and agentic loops where the model makes many short calls rather than one enormous one. Speed and price were the point — M2.1 lists at $0.30 per million input tokens against roughly $0.55 for M1 at independent providers. The window was what paid for it. M3 gets the million back with a different mechanism again: MiniMax Sparse Attention, on a much larger 427-billion-parameter mixture. The practical lesson for anyone reading a catalogue: a context window belongs to a model, not to a brand. Four consecutive releases from a company famous for long context shipped with a window that would not hold a long novel, and nothing in the version numbers said so.