MiniMax-M1

What it is

MiniMax-M1 is a reasoning model released by MiniMax in June 2025. A reasoning model is one trained to work through a problem in steps before giving its answer, and M1 was offered as an open model whose weights were published for others to use.^[2] It was released in two versions, with 40K and 80K thinking budgets, named MiniMax-M1-40k and MiniMax-M1-80k. The thinking budget is the amount of reasoning the model is allowed to produce before it must answer.^[1][3]

The weights were posted to Hugging Face, the public platform where researchers and companies share models.^[1]

Origins

M1 followed the MiniMax-01 series, which had been released as open weights in January 2025. It uses the same kind of hybrid design, with lightning attention and mixture-of-experts layers.^[1][2]

M1 was in turn followed by MiniMax-M2, whose weights were posted to Hugging Face in October 2025.^[3]

Main features

M1 is a hybrid mixture-of-experts model with lightning attention. It has 456 billion total parameters, and 45.9 billion are activated for each token.^[1] Mixture-of-experts means that only some of the model's specialist sub-networks run on any given input, so the active part of the model is much smaller than its total size. Lightning attention is a linear-attention design that is cheaper to run on long inputs than standard attention.

The model natively supports a context of 1 million tokens, which is the amount of text it can consider at once.^[1] A token is a unit of text, such as a word or part of a word.

At a generation length of 100K tokens it uses 25% of the floating-point operations (FLOPs) that DeepSeek R1 uses for the same length, according to the figures published for the model.^[1] FLOPs measure the amount of arithmetic a model performs, so a lower share means less computation for the same amount of output.

M1 was trained with CISPO, a reinforcement-learning algorithm that clips the importance-sampling weights rather than clipping token updates.^[1] Reinforcement learning adjusts a model using feedback on its outputs, and the choice of which quantities to clip affects how stable that training is.

The 80K version reported the following scores: 86.0 on AIME 2024, 76.9 on AIME 2025, 56.0 on SWE-bench Verified and 65.0 on LiveCodeBench.^[1] On agent and long-context tests it scored 62.0 on TAU-bench airline and 63.5 on TAU-bench retail, 73.4 on OpenAI-MRCR at 128K and 56.2 at 1M.^[1] These figures are the company's published results and are best read alongside the test conditions for each benchmark.

M1 is released under the Apache 2.0 license.^[1]

Place in development

M1 is the reasoning model that MiniMax released in June 2025, after the MiniMax-01 series and before MiniMax-M2.^[2][3] Its importance lies in combining a reasoning-focused design with a very long context and open weights. The later M2 series moved the line toward coding and agent work, which the M2 entry describes in more detail.

For readers, M1 is the model that shows how MiniMax's long-context architecture was adapted for step-by-step reasoning, and it sits between the company's first open series and its later coding-focused models.