MiniMax-M3

Spaces using MiniMaxAI/MiniMax-M3 50

Collection including MiniMaxAI/MiniMax-M3

[

MiniMax-M3

Collection

7 items • Updated 27 days ago • 22

](https://huggingface.co/collections/MiniMaxAI/minimax-m3)

Paper for MiniMaxAI/MiniMax-M3

[

MiniMax Sparse Attention

Paper • 2606.13392 • Published Jun 11 • 166

](https://huggingface.co/papers/2606.13392)

Contact Us

Contact us at model@minimax.io.

Safetensors

Model size

427B params

Tensor type

BF16

·

F32

·

Model tree for MiniMaxAI/MiniMax-M3

Adapters

1 model

Finetunes

13 models

Quantizations

59 models

Local Deployment

Download the model:

hf download MiniMaxAI/MiniMax-M3 --local-dir MiniMax-M3

We recommend the following inference frameworks to serve the model:

• SGLang - see SGLang cookbook.

• vLLM - see vLLM recipes.

• Transformers - see Transformers docs.

• KTransformers - see KTransformers MiniMax-M3 tutorial.

• unsloth - see tutorial

• ATOM - see MiniMax-M3 MXFP4/MXFP8 Usage Guide

Inference Parameters

We recommend the following parameters for best performance: temperature=1.0, top_p=0.95.

How to Use

• MiniMax Agent
• MiniMax API

M3 supports three reasoning modes through the thinking parameter:

• enabled — Reasoning is always enabled.
• adaptive — M3 automatically determines when additional reasoning is beneficial.
• disabled — Reasoning is disabled to minimize latency and maximize throughput.

MiniMax Sparse Attention (MSA)

M3 is powered by MiniMax Sparse Attention (MSA), a high-performance sparse attention operator designed for million-token contexts. Compared with GQA, MSA dramatically reduces the attention compute and memory footprint while preserving model quality.

📄 Read the technical report: arXiv:2606.13392 · Hugging Face Papers

MiniMaxAI/MiniMax-M3 · Hugging Face

We’re on a journey to advance and democratize artificial intelligence through open source and open science.

Page Metadata

• Downloads last month: 186,231
• Like: Like
• Show MiniMax's followers: 10,500

Hugging Face

[

Image-Text-to-Text

](https://huggingface.co/models?pipeline_tag=image-text-to-text)[

Transformers

](https://huggingface.co/models?library=transformers)[

Safetensors

](https://huggingface.co/models?library=safetensors)[

minimax_m3_vl

](https://huggingface.co/models?other=minimax_m3_vl)[

multimodal

](https://huggingface.co/models?other=multimodal)[

Mixture of Experts

](https://huggingface.co/models?other=moe)[

agent

](https://huggingface.co/models?other=agent)[

coding

](https://huggingface.co/models?other=coding)[

video

](https://huggingface.co/models?other=video)[

conversational

](https://huggingface.co/models?other=conversational)[

custom_code

](https://huggingface.co/models?other=custom_code)[

Eval Results

](https://huggingface.co/models?other=eval-results)

Model card [Files Files and versions

xet

](https://huggingface.co/MiniMaxAI/MiniMax-M3/tree/main)[Community

29

](https://huggingface.co/MiniMaxAI/MiniMax-M3/discussions)

MiniMax-M3 is a native multimodal model with 1M context. It has ~428B parameters and ~23B activated parameters.

Highlights:

• Native Multimodality: M3 undergoes mixed-modality training from the very first step, enabling deeper semantic fusion across text, image, and video.
• Context Scaling via Sparse Attention: M3 introduces MiniMax Sparse Attention (MSA) to improve long context efficiency. M3 delivers 9× prefill and 15× decode speedups compared to M2 at 1M context, reducing per-token compute to 1/20.
• Coding & Cowork Capability: M3 achieves frontier-level performance across long-horizon agentic benchmarks, excelling in both coding and cowork.

![](https://huggingface.co/MiniMaxAI/MiniMax-M3/resolve/main/figures/ben

(section continues in the model card)