MiniMax M3

Parameters

428B

Context Length

1M

Modality

Multimodal

Architecture

Dense

KV Cache Rate

120 KB/tok

License

Proprietary

Release Date

1 Jun 2026

Knowledge Cutoff

-

API Pricing (per 1M)

Input: $0.30 · Output: $1.20

Hardware & System Requirements

Precision

Native

Context

Workload Memory: 843.49 GB

FP16 · 8K context

Base Weights: 840.05 GB

KV Cache: 0.94 GB

Runtime & Graphs: 2.50 GB

Minimum Required Hardware

8x NVIDIA H200 141GB

843.5 / 1015.2 GB usable · TP 8

Determine requirements for your specific workload:

Upgrade to Pro

Runtime requirements

Checkpoint format

FP16

CUDA

>= 11.8

PyTorch

>= 2.1

Transformers

>= 4.40.1

ArchitecturePrecisionAttention kernel

Blackwell, Hopper

Native FP16

FlashAttention-3

Ada Lovelace, Ampere

Native FP16

FlashAttention-2

Architecture Diagram

Input TokensToken EmbeddingPosition: AbsoluteHidden: 6.1k · Context: 1M · Vocab: 200.1kx 60 layersRMSNormPre-AttentionMulti-Head Attention64Q / 4KV headsHead dim: 128+RMSNormPre-FFNFeed-Forward NetworkSwiGLUIntermediate: 12.3k+Final RMSNormOutput Logits

Evaluation Benchmarks

Rank

#145

BenchmarkScoreRank

0.825

13

0.76

32

Agentic Coding

LiveBench Agentic

0.41

54

0.67

56

0.77

56

LiveBench Average

LiveBench Average

0.67

56

Agent Arena

Agent Arena

959.55

56

0.74

58

Agentic Index

Artificial Analysis

0.29

59

0.68

60

0.59

69

Web Development

WebDev Arena

1484.35

71

General Text

Text Arena

1434.37

96

Intelligence Index

Artificial Analysis

0.19

189

Rankings

Overall Rank

#145

Coding Rank

#99

About MiniMax M3

MiniMax's flagship multimodal model released June 1, 2026. Powered by MiniMax Sparse Attention (MSA) architecture, which replaces traditional full attention with a KV-block selection pattern, drastically reducing compute costs to 1/20th of the previous generation. It is highly optimized for long-horizon agentic workflows, complex software engineering, and video understanding. Features a 1M token context window, supports text, image, and video inputs, and is priced at $0.30 per million input tokens and $1.20 per million output tokens.

Specifications

Attention

Attention Structure

Multi-Head Attention

Attention Heads

64

Key-Value Heads

4

Attention Head Dimension

128

Position Embedding

Absolute Position Embedding

RoPE Theta

5,000,000

Normalization

RMS Normalization

Activation Function

SwigLU

Dimensions

Base Precision

FP16

Hidden Dimension Size

6,144

Number of Layers

60

FFN Intermediate Size (Dense)

12,288

Multi-Token Prediction Heads

1

Tokenizer

Vocabulary Size

200,064

About MiniMax M3

MiniMax's flagship M3 model family, released June 1, 2026, is powered by MiniMax Sparse Attention (MSA) architecture, offering 1M context capabilities at exceptionally low compute cost and optimized for long-horizon agentic workflows.


Other MiniMax M3 Models