MiniMax-Text-01

MiniMax-Text-01, published on 15 January 2025, is the model on which the Shanghai lab built its reputation and every model it has shipped since. It is a mixture of experts totalling 456 billion parameters across 80 layers, routing two of 32 experts per token so that 45.9 billion parameters are active on any given token — a ratio the company has revisited in every later generation. Its distinguishing idea is the attention stack. Rather than use softmax attention throughout, MiniMax interleaves it with lightning attention, a linear-cost variant, placing one softmax layer after every seven lightning layers. Full attention grows quadratically with sequence length, which is what makes long windows expensive; the hybrid keeps most layers linear and reserves the expensive mechanism for one layer in eight. The result, according to MiniMax, is a model trained at a one-million-token context that can be run at up to four million tokens at inference. For January 2025 that was an order of magnitude beyond what open-weight competitors offered. The weights carry MiniMax’s own model agreement rather than an OSI-approved licence — the accompanying code is MIT, the weights are not. The model remains available: independent providers list it as minimax-01 at roughly $0.20 per million input tokens and $1.10 per million output, and it is still the architectural parent of MiniMax-M1.

▸Videos

No videos yet.

▸Reviews