MiniMax-M1: Scaling Test-Time Compute Efficiently with Lightning Attention (arXiv:2506.13585)

Fields

id

sources/model-cn-minimax-m1-paper

URL

https://arxiv.org/abs/2506.13585

Publisher

arXiv / MiniMax

Kind

whitepaper

Published

2025-06-16

Accessed

2026-10-03

notes

Primary source for MiniMax-M1's hybrid attention pattern: 456B total / 45.9B activated, combining MoE (32 local experts, 2 per token) with a lightning-attention mechanism interleaved with full attention. The config.json layer_types array confirms the exact split - 70 linear_attention layers and 10 full_attention layers out of 80.

Notes

Primary source for MiniMax-M1's hybrid attention pattern: 456B total / 45.9B activated, combining MoE (32 local experts, 2 per token) with a lightning-attention mechanism interleaved with full attention. The config.json layer_types array confirms the exact split - 70 linear_attention layers and 10 full_attention layers out of 80.

Cross-references

This record references no other records.