MiniMax M3.1 Flash

[reasoning]

MiniMax[text][open-weights]

Multimodal successor to M3 with clear gains on complex agent tasks, coding and cybersecurity, at Flash pricing with a 1M context

Interactive Architecture

Take MiniMax M3.1 Flash apart, layer by layer. The 3D teardown below shows its 60 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 the same 200K-token vocabulary as M3. Image and video tokens from the vision encoder join this stream (not drawn).

60 layers, all sparse attention

M3 kept 3 full-attention layers; M3.1 Flash makes all 60 sparse. The first 3 layers still use a dense 12,288-dim FFN and the remaining 57 are MoE.

MiniMax Sparse Attention everywhere

GQA (64 query / 4 KV heads) with per-head QK-Norm, wrapped in a block-sparse scheme: a lightning indexer with 4 heads scores 128-token blocks and each query reads only the top 16 plus its local block. The KV cache is about 120 KiB per token at bf16. (Rendered as the GQA block.)

128-expert MoE

A sigmoid router with a learned bias picks 4 of 128 experts per token, plus 1 shared expert, each 3072 dims wide. About 23B of 428B parameters fire per token.

Output head

Final RMSNorm then an untied projection to 200,064 logits. A DSpark draft head speeds up decoding by proposing tokens ahead; it is not drawn.

Input

$0.15

per 1M tokens

Output

$0.60

per 1M tokens

Context

1.0M tokens

Max Output

131K tokens

Released

2026-10-18

Features

[optional-reasoning][1m-tokens-context][vision-and-video-input][agentic-coding][open-weights]

Technical Specifications

LicenseMiniMax Community License

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 M3.1 Flash use?+

6144-dim vectors over the same 200K-token vocabulary as M3. Image and video tokens from the vision encoder join this stream (not drawn).

How many transformer layers does MiniMax M3.1 Flash have?+

M3 kept 3 full-attention layers; M3.1 Flash makes all 60 sparse. The first 3 layers still use a dense 12,288-dim FFN and the remaining 57 are MoE.

What attention does MiniMax M3.1 Flash use?+

How does the Mixture of Experts in MiniMax M3.1 Flash work?+

A sigmoid router with a learned bias picks 4 of 128 experts per token, plus 1 shared expert, each 3072 dims wide. About 23B of 428B parameters fire per token.

How does MiniMax M3.1 Flash produce output tokens?+

Final RMSNorm then an untied projection to 200,064 logits. A DSpark draft head speeds up decoding by proposing tokens ahead; it is not drawn.

scene values sourced from:HuggingFace config HuggingFace config https://sebastianraschka.com/llm-architecture-gallery/#card-minimax-m3-1-flash-428b

Model ID: minimax-m3.1-flash

Interactive Architecture

Take MiniMax M3.1 Flash apart, layer by layer. The 3D teardown below shows its 60 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 the same 200K-token vocabulary as M3. Image and video tokens from the vision encoder join this stream (not drawn).

60 layers, all sparse attention

M3 kept 3 full-attention layers; M3.1 Flash makes all 60 sparse. The first 3 layers still use a dense 12,288-dim FFN and the remaining 57 are MoE.

MiniMax Sparse Attention everywhere

128-expert MoE

A sigmoid router with a learned bias picks 4 of 128 experts per token, plus 1 shared expert, each 3072 dims wide. About 23B of 428B parameters fire per token.

Output head

Final RMSNorm then an untied projection to 200,064 logits. A DSpark draft head speeds up decoding by proposing tokens ahead; it is not drawn.

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 M3.1 Flash use?+

6144-dim vectors over the same 200K-token vocabulary as M3. Image and video tokens from the vision encoder join this stream (not drawn).

How many transformer layers does MiniMax M3.1 Flash have?+

M3 kept 3 full-attention layers; M3.1 Flash makes all 60 sparse. The first 3 layers still use a dense 12,288-dim FFN and the remaining 57 are MoE.

What attention does MiniMax M3.1 Flash use?+

How does the Mixture of Experts in MiniMax M3.1 Flash work?+

A sigmoid router with a learned bias picks 4 of 128 experts per token, plus 1 shared expert, each 3072 dims wide. About 23B of 428B parameters fire per token.

How does MiniMax M3.1 Flash produce output tokens?+

Final RMSNorm then an untied projection to 200,064 logits. A DSpark draft head speeds up decoding by proposing tokens ahead; it is not drawn.