Mistral-Large-3-675B-Base-2512
Mistral AI · base checkpoint
Open in the comparison table Hugging Face ↗ Paper ↗
KV calculations use one sequence at 131,072 tokens and 16-bit cache elements. They include bounded sliding-window caches. This is retained cache payload, not peak prefill memory. The comparison table also offers sourced mixed-precision layouts, including quantization scales.
Release
- Released
2025-12-02
Source: blog · Introducing Mistral 3; header
Introducing Mistral 3 December 2, 2025
Same post: "We release both the base and instruction fine-tuned versions of Mistral Large 3".
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.
- License
apache-2.0
Source: model-card · README.md; YAML front matter
license: 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.
- Input modalities
text + image
Source: model-card · README.md; Key Features
Enables the model to analyze images and provide insights based on visual content, in addition to text.
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").
Parameters
- Total params
675B parameters
Source: model-card · README.md; intro
with 41B active parameters and 675B total parameters
Total parameter count as reported by the publisher. See params.note for whether embeddings, MTP heads or vision towers are included.
- Active params
41B parameters
Source: model-card · README.md; intro
with 41B active parameters and 675B total parameters
Parameters used per token in the forward pass, as reported by the publisher.
- Active non-embedding (stated)
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.
- Counting scope
Headline counts include the 2.5B vision encoder; the language model alone is 673B total, 39B active.
Source: model-card · README.md; Key Features
A Granular MoE Language Model with 673B params and 39B active
What the publisher's counts include or exclude (MTP modules, embeddings, vision encoder, n-gram tables).
Shape
- Layers
61
Source: config · /n_layers
Number of transformer blocks in the main backbone (attention or token mixer plus FFN), excluding MTP modules.
- d_model
7,168
Source: config · /dim
Residual stream width.
- Vocab
131,072
Source: config · /vocab_size
Embedding table rows (vocab_size in config, may include padding).
- Tied embeddings
No
Source: config · /tied_embeddings
Input embedding and output head share weights.
- d_model / layers
117.508
Calculated: d_model / layers
Width-to-depth ratio, d_model divided by the number of blocks.
Attention
- Token mixers
MLA
Source: code · vllm models/mistral_large_3.py; MistralLarge3ForCausalLM
class MistralLarge3ForCausalLM(DeepseekV3ForCausalLM):
Checkpoint attention.wkv_a_with_mqa is remapped to DeepSeek-V3 kv_a_proj_with_mqa (512 latent + 64 RoPE key).
Token-mixing layer types in the backbone, e.g. GQA, MLA, SWA, Gated DeltaNet, Mamba-2, DSA (DeepSeek sparse attention).
- Global-attention layers
61
Source: config · /n_layers
Every block has the same MLA attention tensors; params.json defines no sliding window.
Layers whose KV cache grows with full context (dense or sparse global attention, MLA). Excludes sliding-window and linear/recurrent layers.
- Query heads
128
Source: config · /n_heads
Query heads in the global-attention layers.
- Heads × head dim
24,576
Calculated: heads × head_dim
attn.heads × attn.head_dim, the query projection width of global layers.
- QK head dim
192
Source: config · /qk_nope_head_dim + /qk_rope_head_dim
params.json /head_dim is also 192.
Query/key dimension per head in global-attention layers (for MLA, nope + rope dims).
- V head dim
128
Source: config · /v_head_dim
Value dimension per head in global layers (MLA v_head_dim; stated for MLA even when equal to the QK head dim).
- MLA KV latent
512
Source: config · /kv_lora_rank
Width of the compressed KV latent that MLA caches per token per layer.
- MLA Q latent
1,536
Source: config · /q_lora_rank
Width of the low-rank query projection in MLA; 0 when queries use a full-rank projection.
- Decoupled RoPE dim
64
Source: config · /qk_rope_head_dim
Width of the shared positional key that MLA caches next to the latent.
- Positional encoding
partial RoPE (MLA decoupled, 64 of 192 QK dims)
Source: code · vllm layers/mla.py; MultiHeadLatentAttentionWrapper.forward_native
q[..., self.qk_nope_head_dim :], k_pe = self.rotary_emb(
Rotary applied to the qk_rope_head_dim (64) slice only; the 128 nope dims are unrotated.
Positional scheme in global layers: RoPE, partial RoPE, NoPE, or mixed per layer type.
- RoPE base
10,000
Source: config · /rope_theta
RoPE base frequency of the global layers at release.
- QK norm
No
Source: code · vllm layers/mla.py; MultiHeadLatentAttentionWrapper.forward_native
kv_c_normed = self.kv_a_layernorm(kv_c)
RMSNorm only on the Q and KV latents (q_a_norm, kv_a_norm); per-head Q/K are not normalized.
Queries and keys are normalized per head before the dot product.
- Attention scaling
queries × (1 + 0.1·ln(1 + ⌊pos/8192⌋)) (Llama-4 style)
Source: code · vllm models/deepseek_v2.py; _get_llama_4_scaling
scaling = 1 + scaling_beta * torch.log( 1 + torch.floor(positions / original_max_position_embeddings)
beta and length from params.json /llama_4_scaling (0.1, 8,192); added for Mistral Large 3 in the same PR.
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.
- QKV bias
No
Source: config · /weight_map | has bias
No bias tensor anywhere in the checkpoint.
Query/key/value projections have bias terms.
Feed-forward
- Activation
SwiGLU
Source: code · vllm models/deepseek_v2.py; DeepseekV2MLP.__init__
Only silu is supported for now.
SiluAndMul over gate (w1) and up (w3) in dense and expert FFNs; params.json sets no activation and the Mistral adapter defaults it to silu.
FFN activation (SwiGLU, GeGLU, ...).
- Dense FFN dim
16,384
Source: config · /hidden_dim
Used by the 3 leading dense blocks; checkpoint layers.0.feed_forward.w1 is [16384, 7168].
Intermediate width of the dense FFN layers that are actually executed.
- Dense FFN / d_model
2.28571
Calculated: ffn.dim / d_model
ffn.dim divided by shape.d_model.
Mixture of experts
- Routed experts
128
Source: config · /moe/num_experts
Routed experts per MoE layer.
- Experts per token
4
Source: config · /moe/num_experts_per_tok
/moe/num_expert_groups and /moe/num_expert_groups_per_tok are 1: no group-limited routing.
Routed experts selected per token.
- Expert dim
4,096
Source: config · /moe/expert_hidden_dim
Intermediate width of one routed expert.
- Leading dense layers
3
Source: config · /moe/first_k_dense_replace
/moe/route_every_n 1 (moe_layer_freq): every later block is MoE.
Number of initial blocks that use a dense FFN instead of MoE.
- Router score
softmax
Source: code · vllm transformers_utils/configs/mistral.py; _remap_moe_args
config["scoring_func"] = "softmax"
Same function sets topk_method None, so no score-correction bias.
Gating function over expert logits (softmax, sigmoid).
- Load balancing
Not disclosed
Load-balancing method, with coefficients where disclosed.
- Routed scaling factor
1
Source: config · /moe/routed_scale
Multiplier applied to the combined routed-expert output (routed_scaling_factor).
- Renormalize top-k
Yes
Source: code · vllm transformers_utils/configs/mistral.py; _remap_moe_args
config["norm_topk_prob"] = True
Selected expert weights are renormalized to sum to 1.
- Expert dim / d_model
0.571429
Calculated: moe.expert_dim / d_model
moe.expert_dim divided by shape.d_model.
- Sparsity (experts / top-k)
32
Calculated: experts / top_k
Routed experts divided by experts selected per token.
- Active / total
0.0607407
Calculated: params.active / params.total
params.active divided by params.total.
Multi-token prediction
- MTP depth
Not disclosed
Additional future tokens predicted during pretraining (0 when the model was trained without MTP).
Normalization and stability
- Norm
RMSNorm
Source: code · vllm models/deepseek_v2.py; DeepseekV2DecoderLayer.__init__
self.input_layernorm = RMSNorm(config.hidden_size, eps=config.rms_norm_eps)
Normalization layer type in the blocks.
- Norm placement
pre
Source: code · vllm models/deepseek_v2.py; DeepseekV2DecoderLayer.forward
hidden_states, residual = self.input_layernorm(hidden_states, residual)
Checkpoint attention_norm and ffn_norm map to input_layernorm and post_attention_layernorm, applied before attention and FFN.
Pre-norm, post-norm, or sandwich (pre and post).
- Norm epsilon
1e-06
Source: config · /norm_eps
Epsilon of the block normalization layers.
Context length
- Pretraining context
Not disclosed
Sequence length of the main pretraining stage.
- Trained context
Not disclosed
Longest sequence length trained on after context extension.
- Advertised context
262.1K tokens
Source: model-card · README.md; Key Features
Supports a 256k context window.
params.json /max_seq_len is 262144.
Longest context the publisher says the model supports, possibly beyond trained length via RoPE scaling.
- Extension method
YaRN factor 36 over original 8,192 (release params.json)
Source: config · params.json; yarn
"yarn": { "original_max_position_embeddings": 8192, "factor": 36,
Training-time context stages are not disclosed. Same dict: beta 32, alpha 1, apply_scale false; params.json also sets llama_4_scaling (beta 0.1, original 8,192).
How context was extended (stages, RoPE base change, YaRN factor).
Pretraining recipe
- Pretraining tokens
Not disclosed
Tokens seen in pretraining including annealing and context extension, as reported.
- Optimizer
Not disclosed
Optimizer used for the bulk of parameters.
- Peak LR
Not disclosed
Peak learning rate of the main pretraining stage.
- Final LR
Not disclosed
Learning rate at the end of the decay.
- LR schedule
Not disclosed
Learning-rate schedule shape with phase boundaries.
- Warmup
Not disclosed
Warmup length in steps or tokens.
- Batch size
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.
- Weight decay
Not disclosed
Decoupled weight decay coefficient.
- Adam betas
Not disclosed
(beta1, beta2) of the Adam-family optimizer or its AdamW fallback.
- Grad clip
Not disclosed
Global gradient-norm clipping threshold.
- Training precision
Not disclosed
Matmul precision of pretraining (BF16, FP8 recipe, NVFP4 recipe) with the scaling granularity where disclosed. Never inferred from the released checkpoint dtype.
- Data mix
multilingual text and images: Common Crawl and other public datasets, licensed and third-party datasets, synthetic data, user interaction data (subject to opt-out)
Source: paper · §2.1 Publicly available datasets
The datasets used to train the model include Common Crawl.
§1.3 ticks "More than 10 trillions tokens" for the text training data; §2.4-2.5 cover user and synthetic data.
Pretraining data composition as disclosed (domain shares, synthetic fraction, repetition).
- Data cutoff
July 2025
Source: paper · §1.3 Latest date of data acquisition/collection
The latest date of data collection was July 2025.
Pretraining data cutoff date.
- Hardware
3000 H200 GPUs
Source: model-card · README.md; intro
trained from scratch with 3000 H200s
Accelerators used for pretraining, count and duration where stated.
- Accelerators
3,000 H200
Source: model-card · README.md; intro
trained from scratch with 3000 H200s
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").
KV cache
- KV bytes / token
70,272 bytes
Calculated: kv.bytes_at_context / 131072 tokens; uniform BF16, payload only; quantization overhead excluded. Global KV: 61 × (576 × 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.
- Global KV bytes / token
70,272 bytes
Calculated: 70272 B / 1 tokens (Global KV); uniform BF16, payload only; quantization overhead excluded. Global KV: 61 × (576 × 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.
- KV bytes / token / B active
1,714 bytes
Calculated: kv.bytes_per_token (131072 tokens) / (params.active / 1e9); uniform BF16, payload only; quantization overhead excluded. Global KV: 61 × (576 × 16/8) B/entry
kv.bytes_per_token divided by active parameters in billions (131,072 tokens in exports).
- KV at context / B active
224,651,014 bytes
Calculated: kv.bytes_at_context (131072 tokens) / (params.active / 1e9); uniform BF16, payload only; quantization overhead excluded. Global KV: 61 × (576 × 16/8) B/entry
kv.bytes_at_context divided by active parameters in billions (131,072 tokens in exports).
- KV at context
9,210,691,584 bytes
Calculated: 70272 B × 131072 (Global KV); uniform BF16, payload only; quantization overhead excluded. Global KV: 61 × (576 × 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.
Evaluations
- AA Intelligence Index
9
Source: leaderboard · Artificial Analysis model page /models/mistral-large-3, Wayback snapshot 2026-10-19
Mistral Large 3 scores 9 on the Artificial Analysis Intelligence Index
Score of Mistral-Large-3-675B-Instruct-2512, post-trained from this base. Variant: Mistral Large 3. Index v4.3.2.
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).