Robots Atlas>ROBOTS ATLAS
MiniMax-M3
AI Models

MiniMax-M3

M3
Natively multimodal MoE model from MiniMax with sparse attention (MSA), up to 1M-token context and open weights; ~428B parameters (23B active).
โœ“ Activeโœ“ Public accessโš– Open weightsLLMMultimodalReasoning modelTool-using model
Context window
1M
tokens
Parameters
428B (23B active)
parameters
๐ŸขMiniMaxProducer
Access:APIDownloadHostedDeployment:โ˜ Cloud๐Ÿ’ป Local

Overview

MiniMax-M3 is a natively multimodal language model from the Chinese company MiniMax. It uses a Mixture of Experts (MoE) architecture with the proprietary MiniMax Sparse Attention (MSA) mechanism. It accepts text and image input and produces text output.

The model supports a context window of up to 1 million tokens (with a guaranteed minimum of 512K) and offers multiple reasoning modes. Its weights are publicly available on Hugging Face under the minimax-community license; the total parameter count is about 428B, of which roughly 23B are active.

Classification
LLMMultimodalReasoning modelTool-using model
Access & deployment
APIDownloadHosted
CloudLocal
Weights: Open weights
Key parameters
๐Ÿ“ Context: 1M
๐Ÿงฉ Parameters: 428B (23B active)
โœ“ Tools
๐Ÿ“ฅ Input: text, image

Technical specification

Context window
1M
tokens
Parameters
428B (23B active)
parameters
License
minimax-community
Features:โœ“ Tool use
Modalities
โฌ‡ Input
textimage
โฌ† Output
text

Capabilities and applications

Native model capabilities
Long context
Processing very long inputs (tens to hundreds of thousands of tokens) while maintaining coherence.
Category: language
Coding
Generating, completing, explaining and debugging code across multiple programming languages.
Category: coding
Reasoning
The model's ability to perform multi-step logical inference, solve complex problems and decompose tasks into steps.
Category: reasoning
Interleaved Multimodal Input
Category: reasoning
Function Calling
Category: planning

Technical architecture

Sources and related pages