Production model

MiniMax-M1

MiniMax · Released 2025-06-16 · 456B parameters (45.9B active) · 1,000,000 ctx · open weights

Architecture

Positional encoding	RoPE (base 1e7)
Norm placement	Post-Norm
Norm type	RMSNorm
QK-Norm	no
Activation	SwiGLU
Attention	Lightning Attention (7 layers) interleaved with softmax attention (1 layer); ratio 7:1
MoE	32 routed experts (no shared expert); top-2 routing
Other	Inherits MiniMax-Text-01 base architecture (456B / 45.9B / 80 layers / 7:1 hybrid), Reasoning model trained with CISPO reinforcement learning, Released as 40K and 80K 'thinking budget' variants

Techniques used

Lightning Attention — Same 7:1 Lightning + softmax stack as MiniMax-01; paper claims 25% of DeepSeek R1's FLOPs at 100K generation length.
DeepSeekMoE — Same shared + routed expert layout as MiniMax-Text-01 (32 routed + 1 shared, top-2).
Swish-Gated Linear Unit — SwiGLU FFN inherited from MiniMax-Text-01.
Root Mean Square Layer Normalization — Pre-Norm RMSNorm inherited from the MiniMax-Text-01 base architecture.
Rotary Position Embedding — RoPE carried over from MiniMax-Text-01; 1M context with the 7:1 Lightning + softmax hybrid.

MiniMax-M1 (June 2025) is the reasoning-tuned successor to MiniMax-Text-01. The architecture is identical — 456B total parameters, 45.9B active per token, 80 layers, 7:1 hybrid Lightning Attention + softmax stack, 32 routed experts at top-2 routing (no shared expert) — but the post-training pipeline is new.

The headline claim of the paper (arXiv:2506.13585) is that the hybrid attention stack gives M1 a sustained per-token FLOP advantage at long generation lengths:

Compared to DeepSeek R1, M1 consumes 25% of the FLOPs at a generation length of 100K tokens.

This is the FLOP profile that the Lightning Attention design targeted — every fourth or eighth layer pays the quadratic cost of full softmax; the rest scale linearly with sequence length.

The other contribution is CISPO (Clipped Importance Sampling Policy Optimization), a reinforcement-learning algorithm that the team uses to train the reasoning variant within ~$535K of compute on 512 H800 GPUs.

Two checkpoints are released: a 40K-thinking-budget variant and an 80K-thinking-budget variant. Architecturally they are identical; the difference is in the post-training configuration of the maximum reasoning trace length.

For the broader Lightning Attention story see the Lightning Attention entry and the MiniMax-Text-01 spec sheet.

Sources

Export

BibTeX

@article{arxiv_2506_13585,
  title         = {MiniMax-M1},
  author        = {MiniMax},
  year          = {2025},
  eprint        = {2506.13585},
  archivePrefix = {arXiv},
  url           = {https://arxiv.org/abs/2506.13585}
}

CSL JSON

{
  "id": "arxiv_2506_13585",
  "type": "article-journal",
  "title": "MiniMax-M1",
  "author": [
    {
      "literal": "MiniMax"
    }
  ],
  "issued": {
    "date-parts": [
      [
        2025
      ]
    ]
  },
  "URL": "https://arxiv.org/abs/2506.13585",
  "number": "2506.13585",
  "source": "arXiv"
}

RIS

TY  - JOUR
TI  - MiniMax-M1
AU  - MiniMax
PY  - 2025
JO  - arXiv
AN  - arXiv:2506.13585
UR  - https://arxiv.org/abs/2506.13585
ER  -

Models like this