MiniMax M3

MiniMax[text][open-weights]

Open-weight 428B MoE for coding and agentic work with 1M context, native multimodal input, and computer use

Interactive Architecture

Take MiniMax M3 apart, layer by layer. The 3D teardown below shows its 60 transformer layers (3 dense + 57 MoE), GQA attention, and expert routing, with values from the public model config.

60 layers (3 dense → 57 MoE)

GQA attention · 64Q heads

4+1 of 128 experts active

428B params · 23B active

Token embedding

6144-dim vectors over a large 200K-token vocabulary, sized for broad multilingual coverage.

60 layers, mixed dense/sparse MLP

Pre-norm blocks where each layer's MLP can be dense (12,288 intermediate) or sparse MoE (3,072 intermediate) via a per-layer selector, letting MiniMax trade capacity for speed layer by layer.

MiniMax Sparse Attention

GQA (64 query / 4 KV heads) wrapped in a block-sparse scheme: a lightning indexer with 4 index heads and 128-token blocks selects which KV blocks each query attends to, making the 1M context affordable. (Rendered as the GQA block.)

128-expert MoE

A router scores 128 experts and activates 4 per token, spending ~23B of ~428B. The low active count keeps inference cheap relative to the stored knowledge.

Output head

Final RMSNorm then 200K logits. M3 is a native-multimodal model built to use its full 1M context.

Input

$0.30

per 1M tokens

Output

$1.20

per 1M tokens

Context

1.0M tokens

Max Output

512K tokens

Released

2026-05-31

Features

[open-weights][1m-tokens-context][top-open-weight-swe-bench-pro-(59.0%)][text,-image,-and-video-input][computer-use]

Technical Specifications

LicenseMiniMax Community License

Benchmarks

BrowseComp83.5%

MCP Atlas74.2%

MMMU-Pro78.1%

OSWorld-Verified70.1%

SWE-bench Pro59.0%

SWE-bench Verified80.5%

Tau2-bench88.9%

Terminal-Bench66.0%

The Lab

location: Shanghai, Chinafounded: 2022backed by: Alibaba; Tencent; listed on HKEX (100) since January 2026

Founded in early 2022 and headquartered in Shanghai, MiniMax develops foundation models for text (the M series), video (Hailuo and H3), speech and music; it listed on the Hong Kong Stock Exchange in January 2026 with Alibaba and Tencent among its investors. MiniMax M3 (June 2026) is a 428B open-weight MoE with a 1M-token context, and MiniMax H3 (July 2026) is an open multimodal video model. Its consumer products include the Hailuo video app and the Talkie character app.

more about MiniMax →

How It Works

What vocabulary and embedding size does MiniMax M3 use?+

6144-dim vectors over a large 200K-token vocabulary, sized for broad multilingual coverage.

How many transformer layers does MiniMax M3 have?+

Pre-norm blocks where each layer's MLP can be dense (12,288 intermediate) or sparse MoE (3,072 intermediate) via a per-layer selector, letting MiniMax trade capacity for speed layer by layer.

What attention does MiniMax M3 use?+

How does the Mixture of Experts in MiniMax M3 work?+

A router scores 128 experts and activates 4 per token, spending ~23B of ~428B. The low active count keeps inference cheap relative to the stored knowledge.

How does MiniMax M3 produce output tokens?+

Final RMSNorm then 200K logits. M3 is a native-multimodal model built to use its full 1M context.

scene values sourced from:HuggingFace config HuggingFace config

Model ID: minimax-m3

Interactive Architecture

Take MiniMax M3 apart, layer by layer. The 3D teardown below shows its 60 transformer layers (3 dense + 57 MoE), GQA attention, and expert routing, with values from the public model config.

60 layers (3 dense → 57 MoE)

GQA attention · 64Q heads

4+1 of 128 experts active

428B params · 23B active

Token embedding

6144-dim vectors over a large 200K-token vocabulary, sized for broad multilingual coverage.

60 layers, mixed dense/sparse MLP

Pre-norm blocks where each layer's MLP can be dense (12,288 intermediate) or sparse MoE (3,072 intermediate) via a per-layer selector, letting MiniMax trade capacity for speed layer by layer.

MiniMax Sparse Attention

128-expert MoE

A router scores 128 experts and activates 4 per token, spending ~23B of ~428B. The low active count keeps inference cheap relative to the stored knowledge.

Output head

Final RMSNorm then 200K logits. M3 is a native-multimodal model built to use its full 1M context.

The Lab

location: Shanghai, Chinafounded: 2022backed by: Alibaba; Tencent; listed on HKEX (100) since January 2026

more about MiniMax →

How It Works

What vocabulary and embedding size does MiniMax M3 use?+

6144-dim vectors over a large 200K-token vocabulary, sized for broad multilingual coverage.

How many transformer layers does MiniMax M3 have?+

Pre-norm blocks where each layer's MLP can be dense (12,288 intermediate) or sparse MoE (3,072 intermediate) via a per-layer selector, letting MiniMax trade capacity for speed layer by layer.

What attention does MiniMax M3 use?+

How does the Mixture of Experts in MiniMax M3 work?+

A router scores 128 experts and activates 4 per token, spending ~23B of ~428B. The low active count keeps inference cheap relative to the stored knowledge.

How does MiniMax M3 produce output tokens?+

Final RMSNorm then 200K logits. M3 is a native-multimodal model built to use its full 1M context.