MiniMax-M3

MiniMax-M3 is the flagship language model of the AI company MiniMax, released with downloadable weights. Its model card describes it as a native multimodal model with a 1M-token context, about 428 billion parameters and about 23 billion activated parameters, tagged as a mixture-of-experts (MoE) model, one that routes each token through only part of its network Vendor-stated.¹ MiniMax announced it on June 1, 2026,² and the weights went up on Hugging Face on June 2, alongside an MXFP8 copy.³ MiniMax also serves it through its API and its MiniMax Code agent product.⁴

Two design choices set it apart from MiniMax’s M2 series, by the company’s account. It was trained on text, images and video together “from the very first step”, and it uses MiniMax Sparse Attention (MSA), a block-sparse attention built on grouped-query attention.¹⁵ The card claims 9× faster prefill and 15× faster decoding than M2 at 1M context, with per-token compute cut to 1/20 Vendor-stated.¹ The MSA paper, which tests the method on a separate 109-billion-parameter model, reports attention compute per token cut 28.4× at 1M context and 14.2× prefill and 7.6× decoding speedups on H800 GPUs Vendor-stated.⁵ The API supports up to 1M tokens of context, with a guaranteed minimum of 512K, and the model has three reasoning modes: enabled, adaptive and disabled.⁴¹

MiniMax lists M3 at $0.30 input and $1.20 output per 1M tokens for prompts up to 512K tokens, half its earlier rate under a cut its pricing page labels “Permanent 50% off”; above 512K tokens the rates are $0.60 and $2.40 Vendor-stated.⁶ That ties M3 for sixth in Cheapest LLM APIs per 1M Output Tokens (October 2026), as the cheapest top-tier row. Its weights ship under the MiniMax Community License: commercial users must display “Built with MiniMax M3”, and products or services with more than $20 million in yearly revenue need MiniMax’s prior written authorization, terms that Chinese Open-Weight Models: Downloads, Derivatives and Token Share compares with other labs’ licenses.⁷ An appendix to the license also lists prohibited uses, including any military purpose.⁷ Hugging Face counted 197,786 downloads of the M3 repository in the 30 days to Oct 13, 2026 Measured.³