MiniMax-01 (Text-01 and VL-01)

What it is

MiniMax-01 is a model series released by MiniMax in January 2025. It has two members: MiniMax-Text-01, a language model, and MiniMax-VL-01, a vision-language model. Both were released as open weights, meaning the trained parameters were posted for others to download.^[2] The weights for both were posted to Hugging Face on 12 January 2025, and the technical report, arXiv 2501.08313, was published on 14 January 2025.^[3][1]

Hugging Face is the public platform where researchers and companies share model weights and documentation. A technical report is the document in which a model's design and results are described in detail.

Origins

MiniMax-01 followed the abab line of proprietary text models that MiniMax had developed earlier. The change to open weights in January 2025 marked a shift in how the company released its language models.^[2]

The next model in the open line was MiniMax-M1, released in June 2025.^[2] MiniMax-01 is therefore the company's first open-weight model series.

Main features

MiniMax-Text-01 has 456B total parameters, and 45.9B of them are activated for each token processed.^[1] The model uses a hybrid architecture. It places a softmax attention layer after every seven Lightning Attention layers, a linear-attention design, and it combines these layers with mixture-of-experts components.^[1] Softmax attention is the standard attention mechanism in many transformer models, and linear attention is a cheaper alternative whose cost grows more gently with input length.

The model has 80 layers, 64 attention heads and a hidden size of 6,144. Its mixture-of-experts part has 32 experts, and each token is routed to the top two.^[1]

Context length is one of the defining features. Training context extends to 1 million tokens, and inference can handle up to 4 million tokens.^[1] A token is a unit of text such as a word or part of a word, and the context length is the amount of text the model can consider at once.

MiniMax reported several scores for the model. On MMLU it reported 88.5, on MMLU-Pro 75.7, on GPQA Diamond 54.4, on MATH 77.4, on HumanEval 86.9 and on IFEval 89.1.^[1] For long-context tests it reported 0.910 on RULER at 1M tokens and 56.5 on LongBench v2 with chain of thought.^[1] These are the company's reported figures, and they should be read with the test conditions in mind.

The weights are released under the MiniMax Model Agreement, and the code is released under the MIT License.^[1]

Place in development

MiniMax-01 is the open model series that followed the proprietary abab line and preceded the reasoning model MiniMax-M1, which was released in June 2025.^[2] Its significance lies in its long context and its hybrid design.

For readers, MiniMax-01 is the point at which MiniMax's model work became visible to outside researchers through open weights.