MiniMax-M3
Native multimodal model with 1M context and sparse attention
MiniMax-M3 is a native multimodal language model released by MiniMax AI. It features approximately 428 billion parameters and supports a one million token context window. The model uses sparse attention to handle long contexts efficiently across text, image, and video inputs.
- Supports a context window of 1 million tokens
- Uses ~23 billion active parameters out of ~428 billion
- Features MiniMax Sparse Attention for efficiency
- Native training on text, image, and video data
- Offers three reasoning modes: enabled, adaptive, disabled
What it is
MiniMax-M3 is a large language model built for mixed-modality tasks. It processes text, images, and video from the start of its training process. The system contains roughly 428 billion parameters in total. Only about 23 billion are active for any given token. This design allows it to manage a context window of one million tokens. It aims to perform well in coding and collaborative agent tasks.
How it came about
The model follows the M2 version. MiniMax AI developed M3 to improve efficiency and capability over its predecessor. The company introduced MiniMax Sparse Attention to address the computational costs of long contexts. Technical details were shared in a report on arXiv. The model is available for download and via API.
Inside the model
M3 uses a mixture of experts architecture. It activates a small subset of its total parameters during inference. The core innovation is MiniMax Sparse Attention. This operator reduces memory use and compute needs for long sequences. It claims to lower per-token computation to one twentieth of previous methods. Training involves mixed modalities from the first step. This approach helps fuse semantics across different input types like video and text.
What it can do
The model handles long-horizon agentic benchmarks. It shows strong performance in coding and cowork scenarios. Users can control reasoning through a specific parameter. Options include enabled, adaptive, or disabled modes. Adaptive mode lets the model decide when extra thinking is needed. Disabled mode prioritizes speed and throughput. The system supports native processing of images and video alongside text.
How to use it
Weights are available on Hugging Face for local deployment. Recommended frameworks include SGLang, vLLM, and Transformers. The model is also accessible via the MiniMax API and code platform. Inference settings suggest a temperature of 1.0 and top_p of 0.95. Guides exist for using the model with KTransformers, unsloth, and ATOM.
Questions people ask about MiniMax-M3
How many parameters does MiniMax-M3 have?
The model has approximately 428 billion total parameters. However, it uses a mixture of experts design. This means only about 23 billion parameters are activated for each token during processing.
What is MiniMax Sparse Attention?
It is a sparse attention operator designed for million-token contexts. It reduces the computational load and memory footprint compared to standard methods. The company claims it preserves model quality while speeding up prefill and decode stages.
Can I run MiniMax-M3 locally?
Yes. The weights are available for download on Hugging Face. You can serve the model using frameworks like SGLang, vLLM, or Transformers. Specific tutorials are provided for various inference engines.
What types of data can M3 process?
It is a native multimodal model. It accepts text, images, and video as inputs. The training process mixes these modalities from the very first step to enable deeper semantic fusion across them.
Good reading and sources
There is no English Wikipedia article on MiniMax-M3 yet, so the facts on this page come from the sources listed above: huggingface.co/MiniMaxAI/MiniMax-M3. Some are the subject's own pages, so its claims about itself are reported as its claims. EveryLLM wrote the text in its own words, with AI assistance, and compares it with the sources to make sure no passage is copied.
It may contain mistakes, and figures change over time. Nothing here is investment or buying advice. If something is wrong, tell us. Last refreshed 2026-11-14.