MiniMax-M1

2025-06-16

Date the weights (or, if earlier, the technical report) were first made public, YYYY-MM-DD, as dated by the publisher. A Hugging Face repository creation date is used only when no dated publisher page exists, with a note.

apache-2.0

License of the released weights as declared by the publisher (model card metadata or license file), e.g. apache-2.0, mit, or the name of a custom license.

text

Inferred from the source; see the explanation below.

Input modalities of the released checkpoint as the publisher states them, from: text, image, video, audio, joined with ' + ' in that order ("text", "text + image + video").

456B parameters

Total parameter count as reported by the publisher. See params.note for whether embeddings, MTP heads or vision towers are included.

45.9B parameters

Parameters used per token in the forward pass, as reported by the publisher.

Not disclosed

Active parameters excluding the input embedding and the output head (unembedding). Stated by the publisher, or computed from params.active by subtracting vocab × d_model for each of the embedding and output head that params.active is stated to include (basis inferred, note with the arithmetic). Not disclosed when the scope of params.active is not stated.

456.1B parameters

Calculated: embedding + output head + all experts + other weights, MiniMaxAI/MiniMax-M1-80k@8d1494b safetensors headers; reproduces stated total (backbone) and active (non-embedding)

Parameters of the text backbone counted from the released checkpoint's safetensors headers (embedding, output head, all experts and other weights; excluding MTP modules, vision/audio towers and n-gram lookup tables). Shown only when the counts reproduce the publisher's stated total and active parameters within 5% under some scope.

2.458B parameters

Calculated: input embedding + output head, MiniMaxAI/MiniMax-M1-80k@8d1494b safetensors headers

Input embedding plus output head (once if tied), counted from the checkpoint.

453.6B parameters

Calculated: dense + routed backbone weights; excludes input embedding, output head, lookup tables, MTP and other modalities, MiniMaxAI/MiniMax-M1-80k@8d1494b safetensors headers; reproduces stated total (backbone) and active (non-embedding)

Language backbone weights excluding input embeddings, output head, lookup/Engram tables, MTP modules and vision/audio towers. Equals params.counted_total minus params.counted_embedding. Same validation as params.counted_total.

5.688B parameters

Calculated: max(sum of dense and routed weights sharing a full layers.<index> tensor path), all experts included, MiniMaxAI/MiniMax-M1-80k@8d1494b safetensors headers

Largest group of language-backbone weights sharing a full layers.<index> tensor path, including every routed expert. Excludes lookup/Engram tables, embeddings, output head, MTP and other modalities. Hybrid checkpoints may store mixers and FFNs as separate layers. All-gather volume also depends on wrapping, sharding and dtype.

45.94B parameters

Calculated: non-expert weights + routed experts × 2/32, excluding embedding and output head, MiniMaxAI/MiniMax-M1-80k@8d1494b safetensors headers; reproduces stated total (backbone) and active (non-embedding)

Active parameters per token excluding the input embedding and output head, counted from the checkpoint as all non-expert weights plus routed-expert weights × top_k / experts. Same validation as params.counted_total.

Counts are those of MiniMax-Text-01, on which M1 is built; embedding scope not stated.

What the publisher's counts include or exclude (MTP modules, embeddings, vision encoder, n-gram tables).

80

Number of transformer blocks in the main backbone (attention or token mixer plus FFN), excluding MTP modules.

6,144

Residual stream width.

200,064

Embedding table rows (vocab_size in config, may include padding).

No

Input embedding and output head share weights.

byte-level BPE, 200K vocab (MiniMax-Text-01 tokenizer)

Tokenizer type and lineage as described by the publisher (e.g. byte-level BPE shared with an earlier model).

76.8

Calculated: d_model / layers

Width-to-depth ratio, d_model divided by the number of blocks.

Lightning, GQA

Token-mixing layer types in the backbone, e.g. GQA, MLA, SWA, Gated DeltaNet, Mamba-2, DSA (DeepSeek sparse attention).

7 Lightning : 1 GQA

Inferred from the source; see the explanation below.

Token-mixer order, summarized as a pattern or layer positions. The fact's optional sequence lists each mixer in order; a ratio alone does not establish the order. Auxiliary prediction modules and standalone FFNs are excluded.

10

Layers whose KV cache grows with full context (dense or sparse global attention, MLA). Excludes sliding-window and linear/recurrent layers.

7

Calculated: attn.linear_layers / attn.global_layers

attn.linear_layers divided by attn.global_layers (3 for '3 Gated DeltaNet : 1 GQA').

70

Layers with a fixed-size recurrent state (linear attention, DeltaNet variants, Mamba).

64

Query heads in the global-attention layers.

8

Key/value heads in the global-attention layers. Not applicable to MLA, whose cache is a shared latent.

8,192

Calculated: heads × head_dim

attn.heads × attn.head_dim, the query projection width of global layers.

8

Calculated: heads / kv_heads

attn.heads divided by attn.kv_heads in global layers.

128

Query/key dimension per head in global-attention layers (for MLA, nope + rope dims).

partial RoPE (softmax layers); Lightning layers use a fixed per-head decay, no RoPE

Positional scheme in global layers: RoPE, partial RoPE, NoPE, or mixed per layer type.

10,000,000

RoPE base frequency of the global layers at release.

0.5

Fraction of each head's QK dimensions that are rotated.

No

Queries and keys are normalized per head before the dot product.

sigmoid gate on the Lightning-layer output, after the output RMSNorm, before out_proj; none in softmax layers

Gate applied to the attention output before the output projection, with its placement.

No

Query/key/value projections have bias terms.

Lightning attention: SiLU on Q, K, V; per-head exponential decay (ALiBi-style slopes, scaled down linearly with depth); RMSNorm on output; 128×128 state per head

Configuration of the linear / SSM layers (e.g. Gated DeltaNet heads and dims, Mamba-2 state size), as one string.

64

Value heads (or SSM heads) in the linear / SSM layers.

128

Head dimension of the linear / SSM layers (value head dim for DeltaNet variants, head_dim for Mamba-2).

SwiGLU

FFN activation (SwiGLU, GeGLU, ...).

No

FFN projections have bias terms.

32

Routed experts per MoE layer.

2

Routed experts selected per token.

0

Always-on experts per MoE layer.

9,216

Intermediate width of one routed expert.

0

Number of initial blocks that use a dense FFN instead of MoE.

softmax

Gating function over expert logits (softmax, sigmoid).

GShard auxiliary loss (coefficient 0.01 for Text-01, lowered in M1 continued pretraining) + global router synchronizing token counts across EP groups

Load-balancing method, with coefficients where disclosed.

Yes

Selected expert weights are renormalized to sum to 1.

1.5

Calculated: moe.expert_dim / d_model

moe.expert_dim divided by shape.d_model.

16

Calculated: experts / top_k

Routed experts divided by experts selected per token.

0.100658

Calculated: params.active / params.total

params.active divided by params.total.

Not disclosed

Additional future tokens predicted during pretraining (0 when the model was trained without MTP).

RMSNorm

Normalization layer type in the blocks.

post

Pre-norm, post-norm, or sandwich (pre and post).

1e-05

Epsilon of the block normalization layers.

DeepNorm: residual × α = (2N)^0.25 = 3.5566 (N = 80), branch × β = 1

Residual design when it is not a plain sum, e.g. hyper-connections with n streams, gated residuals, depth-scaled residuals, FP32 residual accumulation.

8.192K tokens

Sequence length of the main pretraining stage.

1M tokens

Longest sequence length trained on after context extension.

1M tokens

Longest context the publisher says the model supports, possibly beyond trained length via RoPE scaling.

Text-01: 8K → 128K (RoPE base 5M, 300B tokens) → 512K (10M, 32B) → 1M (10M, 26B); M1 continued pretraining: four stages from 32K to 1M

How context was extended (stages, RoPE base change, YaRN factor).

Not disclosed

Tokens seen in pretraining including annealing and context extension, as reported.

AdamW

Optimizer used for the bulk of parameters.

0.0002

Peak learning rate of the main pretraining stage.

3e-05

Learning rate at the end of the decay.

Text-01: 500-iter warmup, constant 2e-4 for 7.2T tokens, 1.3e-4 for the remaining 3.2T, exponential decay to 3e-5 over 1T; M1: constant 8e-5 for 2.5T, decay to 8e-6 over 5T

Learning-rate schedule shape with phase boundaries.

500 iterations

Warmup length in steps or tokens.

16M → 32M (at 69B tokens) → 64M (790B) → 128M tokens (4.7T), then constant

Global batch size, including any batch-size ramp: sequences × sequence length (token total), or the token total alone when only that is stated.

128M tokens

Global batch size in tokens during the main pretraining stage, after any ramp (sequences × sequence length when stated in sequences, marked inferred).

0.1

Decoupled weight decay coefficient.

(0.9, 0.95)

(beta1, beta2) of the Adam-family optimizer or its AdamW fallback.

Not disclosed

Global gradient-norm clipping threshold.

Not disclosed

Matmul precision of pretraining (BF16, FP8 recipe, NVFP4 recipe) with the scaling granularity where disclosed. Never inferred from the released checkpoint dtype.

Xavier; DeepNorm α = (2N)^0.25, β = (8N)^-0.25

Weight initialization (std or scheme) as stated by the publisher; config initializer_range alone is not evidence.

LR cut from 2e-4 to 1.3e-4 after anomalous gradient norms; M1 context extension spread over four stages after gradient explosions with aggressive extension

Interventions used during pretraining to keep it stable (QK-clip, z-loss, logit caps, spike handling), with coefficients.

Text-01: academic literature, books, web, code, scored by a 5B-active/60B MoE; high-quality data up to 4 epochs. M1: STEM, code, book and reasoning data raised to 70%, no synthetic data

Pretraining data composition as disclosed (domain shares, synthetic fraction, repetition).

1500–2500 H800 GPUs (cluster size varied)

Accelerators used for pretraining, count and duration where stated.

1,500–2,500 H800

Accelerator count and type for the main pretraining run, as the count with thousands separators and the chip's short name ("2,048 H800", "2,048 GH200"). A stated range or bound stays as stated ("2,048–4,096 GH200").

• Text-01 main

  Seq len

  8192

  Lr

  0.0002

  Batch tokens

  128000000

  Source: paper · Text-01 report p.24–25 §4.2, Table 6; M1 report p.3 §2.1

  constant learning rate for 7.2T tokens

  Constant 2e-4 for 7.2T tokens, then 1.3e-4 for 3.2T; batch ramped 16M → 128M by 4.7T tokens.

• Text-01 fast decay

  Tokens

  1000000000000

  Seq len

  8192

  Batch tokens

  128000000

  Source: paper · Text-01 report p.24–25 §4.2, Table 6; M1 report p.3 §2.1

  During the fast decay phase, we train 1T

  Exponential decay to 3e-5.

• Text-01 long-context 128K

  Tokens

  300000000000

  Seq len

  131072

  Source: paper · Text-01 report p.24–25 §4.2, Table 6; M1 report p.3 §2.1

  128K 5M 300B

  RoPE base 5M; 30% short, 70% medium data.

• Text-01 long-context 512K

  Tokens

  32000000000

  Seq len

  524288

  Source: paper · Text-01 report p.24–25 §4.2, Table 6; M1 report p.3 §2.1

  512K 10M 32B

  RoPE base 10M; 35% short, 35% medium, 30% long.

• Text-01 long-context 1M

  Tokens

  26000000000

  Seq len

  1000000

  Source: paper · Text-01 report p.24–25 §4.2, Table 6; M1 report p.3 §2.1

  1M 10M 26B

  RoPE base 10M; 30% short, 30% medium, 40% long.

• M1 continued pretraining

  Tokens

  7500000000000

  Lr

  8e-05

  Source: paper · Text-01 report p.24–25 §4.2, Table 6; M1 report p.3 §2.1

  we continue training the MiniMax-Text-01 model with additional 7.5T tokens

  Constant 8e-5 for 2.5T, then decay to 8e-6 over 5T; context 32K → 1M in four stages; reasoning-heavy mix (70% STEM, code, book, reasoning).

Ordered pretraining stages. Each item: name, tokens, seq_len, lr, batch_tokens, batch_sequences, steps, note, quote, and source when it differs from the fact's source.

Text-01: softmax / lightning / hybrid scaling study at 70M–7B dense (≤300B tokens, 8K); size law fit on 44M–1.2B active MoE (16/32/64 experts, 500B tokens); MoE ablations at 28B-A5B (1T tokens) and 60B-A9.3B, 48 layers (500B tokens)

Smaller models trained to choose the architecture or recipe (sizes, token budgets), as disclosed.

40,960 bytes

Calculated: kv.bytes_at_context / 131072 tokens; uniform BF16, payload only; quantization overhead excluded. Global KV: 10 × (2048 × 16/8) B/entry

Retained KV cache divided by context length, including bounded SWA caches and indexer keys. Exports use uniform BF16 at 131,072 tokens. The website can select uniform payload precision (no quantization overhead) or sourced component formats (including scales and documented padding). This is cache payload, not peak prefill memory or persistent prefix storage.

40,960 bytes

Calculated: 40960 B / 1 tokens (Global KV); uniform BF16, payload only; quantization overhead excluded. Global KV: 10 × (2048 × 16/8) B/entry

Long-context global KV bytes per token, including indexer keys; excludes bounded SWA caches. Compression rates and cross-layer cache reuse are included. Exports use uniform BF16; the website's storage selector also applies here. This is the field to compare with global-KV figures in reports, using the matching storage recipe.

892 bytes

Calculated: kv.bytes_per_token (131072 tokens) / (params.active / 1e9); uniform BF16, payload only; quantization overhead excluded. Global KV: 10 × (2048 × 16/8) B/entry

kv.bytes_per_token divided by active parameters in billions (131,072 tokens in exports).

116,965,340 bytes

Calculated: kv.bytes_at_context (131072 tokens) / (params.active / 1e9); uniform BF16, payload only; quantization overhead excluded. Global KV: 10 × (2048 × 16/8) B/entry

kv.bytes_at_context divided by active parameters in billions (131,072 tokens in exports).

5,368,709,120 bytes

Calculated: 40960 B × 131072 (Global KV); uniform BF16, payload only; quantization overhead excluded. Global KV: 10 × (2048 × 16/8) B/entry

Retained KV cache for one sequence at the selected context, including SWA windows and indexer keys. Exports use uniform BF16 at 131,072 tokens. Excludes recurrent state, compressor tail state, draft/MTP caches, temporary prefill tensors, allocator metadata and persistent prefix snapshots. Does not estimate peak prefill or total serving memory.

12

Artificial Analysis Intelligence Index v4.3.2 of the post-trained model, from a dated Wayback snapshot of its Artificial Analysis page. The note names the variant (reasoning setting).