Mistral-Large-4-1T-A52B

2026-10-29

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.

other (Mistral Ecosystem Open-Weight License)

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 + image

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").

1.05T parameters

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

52B parameters

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

49B parameters

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.

1.049T parameters

Calculated: embedding + output head + all experts + other weights, mistralai/Mistral-Large-4-1T-A52B@f12dabd safetensors headers; reproduces stated total (backbone) and active (with embedding and output head)

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.

2.517B parameters

Calculated: input embedding + output head, mistralai/Mistral-Large-4-1T-A52B@f12dabd safetensors headers

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

1.047T parameters

Calculated: dense + routed backbone weights; excludes input embedding, output head, lookup tables, MTP and other modalities, mistralai/Mistral-Large-4-1T-A52B@f12dabd safetensors headers; reproduces stated total (backbone) and active (with embedding and output head)

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.

14.75B parameters

Calculated: max(sum of dense and routed weights sharing a full layers.<index> tensor path), all experts included, mistralai/Mistral-Large-4-1T-A52B@f12dabd 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.

49.82B parameters

Calculated: non-expert weights + routed experts × 4/128, excluding embedding and output head, mistralai/Mistral-Large-4-1T-A52B@f12dabd safetensors headers; reproduces stated total (backbone) and active (with embedding and output head)

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.

Card: 49B active without and 52B with embeddings and output layers. Docs: 1.05T total "and a 1.6B vision encoder"; whether the total includes the vision encoder or the DSpark draft head is not stated.

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

72

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

6,144

Residual stream width.

204,800

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

No

Input embedding and output head share weights.

Tekken v15 (tekken.json), 204,800 tokens including 1,000 special tokens

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

85.3333

Calculated: d_model / layers

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

MLA, DSA

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

72 DSA; layers 39, 43, …, 71 reuse layer 35's KV cache

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.

72

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

64

Query heads in the global-attention layers.

8,192

Calculated: heads × head_dim

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

128

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

128

Value dimension per head in global layers (MLA v_head_dim; stated for MLA even when equal to the QK head dim).

512

Width of the compressed KV latent that MLA caches per token per layer.

0

Width of the low-rank query projection in MLA; 0 when queries use a full-rank projection.

64

Width of the shared positional key that MLA caches next to the latent.

mixed: partial RoPE (MLA decoupled, interleaved, LeRoPE, 64 of 128 QK dims) in 54 layers, NoPE in 18 (every 4th)

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

10,000

RoPE base frequency of the global layers at release.

No

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

DSA indexer (32 heads × 128, top-2048); 44 layers run their own indexer, 28 reuse the previous indexer layer's top-k

Token-selection mechanism inside global layers, e.g. DSA lightning indexer with top-k.

learned per-head query scaling × (s_h·ln(pos+1) + b_h) (SSMax-style), in attention and the DSA indexer

Query-key scaling when it is not 1/sqrt(head_dim), e.g. query_pre_attn_scalar, length-dependent (SSMax-style) scaling, or value scaling.

No

Query/key/value projections have bias terms.

128

Width of the sparse-attention indexer key cached per token in each global layer (DSA lightning indexer), counted in the KV cache.

9 NoPE layers (39, 43, …, 71) reuse layer 35's latent KV cache and have no kv_a projection; the other 63 layers own caches

Layers that reuse another layer's KV cache instead of computing their own (YOCO-style, cross-layer attention), and which layers own caches.

• MLA latent + RoPE key

  Layers

  54

  Elements

  576

  Tokens per entry

  1

  Source: code · modeling_mistral_large4.py; MistralLarge4TextAttention.__init__, MistralLarge4TextAttention.forward

  kv_pass, k_rot = torch.split(compressed_kv, [self.kv_lora_rank, self.qk_rope_head_dim], dim=-1)

  RoPE layers: kv_lora_rank 512 + qk_rope_head_dim 64.

• MLA latent (NoPE layers)

  Layers

  9

  Elements

  512

  Tokens per entry

  1

  Source: code · modeling_mistral_large4.py; MistralLarge4TextAttention.__init__, MistralLarge4TextAttention.forward

  self.qk_rope_head_dim = config.qk_rope_head_dim if self.use_rope else 0

  NoPE cache owners 3, 7, …, 35; layers 39–71 (every 4th) read layer 35's cache.

• DSA indexer key

  Layers

  44

  Elements

  128

  Tokens per entry

  1

  Source: code · modeling_mistral_large4.py; MistralLarge4TextAttention.__init__, MistralLarge4TextAttention.forward

  self.indexer = None if self.skip_topk else MistralLarge4TextIndexer(config, layer_idx)

  Only "full" indexer layers own an indexer; 28 "shared" layers reuse the previous top-k.

Cache owners and entry shapes for compressed or shared attention. Each item specifies layers owning the cache, scalar elements per entry (K and V together), tokens_per_entry (compression ratio, default 1), and optionally window (maximum entries retained). Completed compressed entries are floor(context / tokens_per_entry); incomplete compressor state is excluded. KV sizes use these groups when present.

SwiGLU

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

24,576

Intermediate width of the dense FFN layers that are actually executed.

No

FFN projections have bias terms.

4

Calculated: ffn.dim / d_model

ffn.dim divided by shape.d_model.

128

Routed experts per MoE layer.

4

Routed experts selected per token.

1

Always-on experts per MoE layer.

6,144

Intermediate width of one routed expert.

6,144

Inferred from the source; see the explanation below.

Total intermediate width of the shared expert(s).

1

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

softmax

Gating function over expert logits (softmax, sigmoid).

Not disclosed

Load-balancing method, with coefficients where disclosed.

1

Multiplier applied to the combined routed-expert output (routed_scaling_factor).

Yes

Selected expert weights are renormalized to sum to 1.

none

Expert groups and groups selected per token, or node-limited routing, when used.

1

Calculated: moe.expert_dim / d_model

moe.expert_dim divided by shape.d_model.

32

Calculated: experts / top_k

Routed experts divided by experts selected per token.

0.0495238

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.

pre

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

30

tanh soft-capping value on output logits.

1e-05

Epsilon of the block normalization layers.

Not disclosed

Sequence length of the main pretraining stage.

Not disclosed

Longest sequence length trained on after context extension.

1.049M tokens

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

YaRN factor 2048 over original 512, with LeRoPE learned per-frequency multipliers (release config)

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

Not disclosed

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

Not disclosed

Optimizer used for the bulk of parameters.

Not disclosed

Peak learning rate of the main pretraining stage.

Not disclosed

Learning rate at the end of the decay.

Not disclosed

Learning-rate schedule shape with phase boundaries.

Not disclosed

Warmup length in steps or tokens.

Not disclosed

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

Not disclosed

Decoupled weight decay coefficient.

Not disclosed

(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.

multilingual text and images: Common Crawl and other public datasets, licensed and third-party datasets, synthetic data, user interaction data from Mistral products (subject to opt-out)

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

September 2026

Pretraining data cutoff date.

3,800 NVIDIA Grace Blackwell GPUs in Mistral's own datacenters in Europe

Accelerators used for pretraining, count and duration where stated.

3,800 Grace Blackwell

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").

82,688 bytes

Calculated: kv.bytes_at_context / 131072 tokens; uniform BF16, payload only; quantization overhead excluded. MLA latent + RoPE key: 54 × (576 × 16/8) B/entry; MLA latent (NoPE layers): 9 × (512 × 16/8) B/entry; DSA indexer key: 44 × (128 × 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.

82,688 bytes

Calculated: 62208 B / 1 tokens (MLA latent + RoPE key) + 9216 B / 1 tokens (MLA latent (NoPE layers)) + 11264 B / 1 tokens (DSA indexer key); uniform BF16, payload only; quantization overhead excluded. MLA latent + RoPE key: 54 × (576 × 16/8) B/entry; MLA latent (NoPE layers): 9 × (512 × 16/8) B/entry; DSA indexer key: 44 × (128 × 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.

1,590 bytes

Calculated: kv.bytes_per_token (131072 tokens) / (params.active / 1e9); uniform BF16, payload only; quantization overhead excluded. MLA latent + RoPE key: 54 × (576 × 16/8) B/entry; MLA latent (NoPE layers): 9 × (512 × 16/8) B/entry; DSA indexer key: 44 × (128 × 16/8) B/entry

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

208,424,645 bytes

Calculated: kv.bytes_at_context (131072 tokens) / (params.active / 1e9); uniform BF16, payload only; quantization overhead excluded. MLA latent + RoPE key: 54 × (576 × 16/8) B/entry; MLA latent (NoPE layers): 9 × (512 × 16/8) B/entry; DSA indexer key: 44 × (128 × 16/8) B/entry

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

10,838,081,536 bytes

Calculated: 62208 B × 131072 (MLA latent + RoPE key) + 9216 B × 131072 (MLA latent (NoPE layers)) + 11264 B × 131072 (DSA indexer key); uniform BF16, payload only; quantization overhead excluded. MLA latent + RoPE key: 54 × (576 × 16/8) B/entry; MLA latent (NoPE layers): 9 × (512 × 16/8) B/entry; DSA indexer key: 44 × (128 × 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.