MiniMax releases MiniMax-M1, world's first open-weight large-scale hybrid-attention reasoning model

MiniMax, the Chinese lab, released MiniMax-M1 with weights published on Hugging Face and GitHub. The company described it as the first open-weight large-scale model to combine a hybrid-attention architecture with reasoning training, and it shipped in two variants — M1-40k and M1-80k — named for the length of “thinking” budget each is tuned to use before answering.

The model is a mixture-of-experts design with 456 billion total parameters, of which 45.9 billion are active for any given token, and it supports a context window of one million tokens natively — eight times the window MiniMax cited for DeepSeek’s R1 at the time. The lightning-attention mechanism underlying it was pitched as a way to make long reasoning chains and long-context inference cheaper at test time than standard transformer attention allows, a claim the accompanying paper set out in detail via reinforcement-learning training on mathematics and software-engineering tasks.

“Open-weight” here means the parameters are downloadable, not that the release carries a standard permissive licence: MiniMax published the model under its own model licence agreement rather than MIT or Apache 2.0, a distinction that mattered to the open-source debate the release fed into. It arrived amid a run of open-weight releases from Chinese labs — following DeepSeek’s R1 earlier in the year and preceding Baidu’s ERNIE 4.5 the same month — that collectively shifted where the frontier of freely downloadable reasoning models sat, even as US export controls continued to constrain the chips those labs could train on.