MiniMax-M3

MiniMaxAI

Parameters

428.0B total / 23.0B active

MoE: total / active

Architecture

Native multimodal MoE with MiniMax Sparse Attention (MSA)

Released

02.06.2026

License

MiniMax Community License

Open Weights Commercial Use Multimodal BF16, F32 MiniMax en zh multilingual

Input Modalities

text image video

Output Modalities

text

Context (native)

1,000,000 tokens

Context (extended)

1,000,000 tokens

Openness Index Score 70.0/100

About

MiniMax-M3 (MiniMaxAI/MiniMax-M3) is a native multimodal Mixture-of-Experts model with a one-million-token context - ~428B total parameters, ~23B activated per token - handling text, image and video input, released June 2, 2026. It undergoes mixed-modality training from the very first step (early fusion), enabling deeper semantic fusion across text, image and video rather than bolting vision onto a text model.

For long-context efficiency M3 introduces MiniMax Sparse Attention (MSA), a high-performance sparse attention operator for million-token contexts that dramatically reduces attention compute and memory footprint versus GQA while preserving quality - at 1M context M3 delivers 9x prefill and 15x decode speedups over M2, cutting per-token compute to 1/20. M3 targets coding and cowork capability, achieving frontier-level performance across long-horizon agentic benchmarks. Technical report: arXiv 2606.13392.

Training Data Mixed-modality training from the first step across text, image, and video

Benchmark Scores

Benchmark Score Date Standing
SWE-bench Verified
coding_agent
91.86%
11.06.2026
#10/60
better than 85%
Terminal Bench 2.1
coding_agent
72.73%
11.06.2026
#25/44
better than 44%
SWE Atlas - QnA
coding_agent
52.37%
11.06.2026
#12/24
better than 52%
SWE Atlas - TW
coding_agent
41.06%
11.06.2026
#12/20
better than 42%
SWE-efficiency
coding_agent
63.80%
11.06.2026
#3/5
better than 50%
LiveSQLBench
coding_agent
89.74%
11.06.2026
#2/5
better than 50%
CL-bench
coding_agent
51.00%
11.06.2026
#4/5
better than 25%
VIBE-V2
coding_agent
79.21%
11.06.2026
#3/9
better than 75%
SVG-Bench
coding_agent
97.52%
11.06.2026
#2/9
better than 88%
PostTrainBench
coding_agent
81.91%
11.06.2026
#7/22
better than 71%

Showing 10 of 41 results

What is a good SWE-bench Verified score? What is a good Terminal Bench 2.1 score? What is a good SWE Atlas - QnA score?

Hardware Requirements

Estimated VRAM per quantization at 4,096 tokens context — MoE: all experts must fit in VRAM, only 23.0B are active per token

Quantization Weights KV Cache Total VRAM On your hardware
FP16 — 16-bit 856.0 GB 3.6 GB 902.4 GB Which GPUs can run it? Calculate on your GPU
INT8 — 8-bit 428.0 GB 1.8 GB 451.2 GB Which GPUs can run it? Calculate on your GPU
INT4 — 4-bit 214.0 GB 0.9 GB 225.6 GB Which GPUs can run it? Calculate on your GPU
FP4 — 4-bit 214.0 GB 0.9 GB 225.6 GB Which GPUs can run it? Calculate on your GPU

Only datacenter GPUs — FP16 needs 902.4 GB. For consumer hardware, choose the INT4 build with a multi-GPU setup.

Architecture

Decoder Block input Embedding vocab 200K · d 6144 Full Attention GQA 64:4 · dₕ 128 ×60 MoE FFN 128 experts · top-4 · +1 shared · dᴻ 3072 MTP Head ×1 speculative layer Final Norm LM Head vocab 200K output
Attention
Grouped Query Attention (64:4)
MoE
128 experts · top-4 per token
Layers
60
Hidden size
6144
Context
1M tokens
RoPE θ
5M
Parameters
428000M
Active params
23000M

Source: Hugging Face config.json · MiniMaxM3SparseForConditionalGeneration · model repo

Type: Mixture of Experts (MoE)
Attention: MiniMax Sparse Attention (MSA)
Decoder: Transformer
MoE: yes (? experts)
Routing: Sparse attention with GQA baseline replacement
Key Innovation

MiniMax Sparse Attention (MSA) for million-token contexts

Training Pipeline

  1. 1
    pretraining

    Mixed-modality pretraining

    Native multimodal pretraining from the very first step, enabling deeper semantic fusion across text, image, and video.

  2. 2
    sft

    Supervised fine-tuning and instruction tuning

    SFT and instruction tuning with support for three reasoning modes (enabled/adaptive/disabled) via the thinking parameter.

Training & Evaluation Datasets

NameRoleSizeModalitiesCollection
Mixed-modality pre-training corpus (text, image, video) pretraining — —

Linked Resources

arxiv

MiniMax Sparse Attention (arXiv:2606.13392)

https://arxiv.org/abs/2606.13392

 github

MiniMax-M3 GitHub Repository

https://github.com/MiniMax-AI/MiniMax-M3

 github

MiniMax Sparse Attention (MSA) GitHub

https://github.com/MiniMax-AI/MSA

 website

MiniMax Agent Demo

https://agent.minimax.io/

 api_docs

MiniMax API Documentation

https://platform.minimax.io/

 website

MiniMax Official Website

https://www.minimax.io/

Comments (0)

No comments yet

Be the first to share your thoughts!

What now?

You have read the full card for MiniMax-M3 — here is where to go next.

Compare with another model Test on your own GPU Related models Data as JSON