DeepSeek-V3-Base

2024-12-26

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 (DeepSeek License Agreement)

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").

671B parameters

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

37B parameters

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

35.7B parameters

Inferred from the source; see the explanation below.

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.

671B parameters

Calculated: embedding + output head + all experts + other weights, deepseek-ai/DeepSeek-V3-Base@afb92e1 safetensors headers; reproduces stated total (backbone) and active (with 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.

1.853B parameters

Calculated: input embedding + output head, deepseek-ai/DeepSeek-V3-Base@afb92e1 safetensors headers

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

669.2B parameters

Calculated: dense + routed backbone weights; excludes input embedding, output head, lookup tables, MTP and other modalities, deepseek-ai/DeepSeek-V3-Base@afb92e1 safetensors headers; reproduces stated total (backbone) and active (with 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.

11.51B parameters

Calculated: max(sum of dense and routed weights sharing a full layers.<index> tensor path), all experts included, deepseek-ai/DeepSeek-V3-Base@afb92e1 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.

35.7B parameters

Calculated: non-expert weights + routed experts × 8/256, excluding embedding and output head, deepseek-ai/DeepSeek-V3-Base@afb92e1 safetensors headers; reproduces stated total (backbone) and active (with 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.

671B excludes the MTP module; the released checkpoint is 685B (671B main model + 14B MTP).

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

61

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

7,168

Residual stream width.

129,280

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

No

Input embedding and output head share weights.

byte-level BPE, 128K vocab; pretokenizer merges punctuation with line breaks, a proportion of these split at random in training

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

117.508

Calculated: d_model / layers

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

MLA

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

61 MLA

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.

61

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

128

Query heads in the global-attention layers.

24,576

Calculated: heads × head_dim

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

192

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.

1,536

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.

partial RoPE (decoupled 64-dim RoPE key shared across heads + 128 NoPE dims per head)

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.

1/sqrt(192) × (0.1·ln 40 + 1)² (YaRN mscale, s = 40)

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.

SwiGLU

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

18,432

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

No

FFN projections have bias terms.

2.57143

Calculated: ffn.dim / d_model

ffn.dim divided by shape.d_model.

256

Routed experts per MoE layer.

8

Routed experts selected per token.

1

Always-on experts per MoE layer.

2,048

Intermediate width of one routed expert.

2,048

Total intermediate width of the shared expert(s).

3

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

sigmoid

Gating function over expert logits (softmax, sigmoid).

aux-loss-free bias (γ=0.001, frozen for the last 500B tokens) + sequence-wise loss α=1e-4

Load-balancing method, with coefficients where disclosed.

2.5

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

Yes

Selected expert weights are renormalized to sum to 1.

node-limited: each token sent to at most 4 of 8 nodes (groups)

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

0.285714

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.

0.0551416

Calculated: params.active / params.total

params.active divided by params.total.

1

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

1

MTP modules with distinct weights. 1 with mtp.depth > 1 means one module is applied recursively.

h' = M_k[RMSNorm(h); RMSNorm(Emb(t_{i+k}))] (2d→d projection), then one Transformer block identical to the main layers (MLA + MoE)

Contents of one MTP module (mixer and FFN type, norms, how the previous hidden state and next-token embedding are combined).

MoE

FFN in the MTP module: 'dense' or 'MoE', with its width when it differs from the backbone.

Yes

MTP modules use the main model's input embedding and output head.

11.5B parameters

Parameters of all MTP modules as reported by the publisher.

0.3 for the first 10T tokens, 0.1 for the remaining 4.8T

Weight of the MTP loss, including its schedule.

Yes

The released checkpoint contains the MTP module weights.

RMSNorm

Normalization layer type in the blocks.

pre

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

1e-06

Epsilon of the block normalization layers.

4.096K tokens

Sequence length of the main pretraining stage.

131.1K tokens

Longest sequence length trained on after context extension.

131.1K tokens

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

YaRN (s=40, α=1, β=32) on the decoupled key; 4K → 32K (1000 steps, batch 1920) → 128K (1000 steps, batch 480), LR 7.3e-6

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

14.8T tokens

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

AdamW

Optimizer used for the bulk of parameters.

0.00022

Peak learning rate of the main pretraining stage.

7.3e-06

Learning rate at the end of the decay.

2K-step warmup, constant to 10T, cosine to 2.2e-5 over 4.3T, then 2.2e-5 for 333B and 7.3e-6 for 167B

Learning-rate schedule shape with phase boundaries.

2000 steps

Warmup length in steps or tokens.

3,072 → 15,360 × 4,096 (62.9M tokens) over the first 469B tokens, 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.

62.91M tokens

Inferred from the source; see the explanation below.

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.

1

Global gradient-norm clipping threshold.

FP8 (E4M3) for Fprop, Dgrad and Wgrad of Linear layers; scales per 1×128 activation tile and 128×128 weight block; embedding, output head, MoE gating, norms and attention kept in BF16/FP32

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

random init, std 0.006 for all learnable parameters

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

packed documents, no cross-sample attention masking

Whether attention crosses document boundaries within a packed sequence.

FIM at rate 0.1, PSM format, applied per document before packing

Objectives beyond next-token prediction and MTP (fill-in-the-middle rate, distillation, multimodal tokens from the start).

14.8T tokens; math and code share raised vs DeepSeek-V2; multilingual coverage beyond English and Chinese

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

2048 H800 GPUs

Accelerators used for pretraining, count and duration where stated.

2,048 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").

PP16 · EP64 · ZeRO-1 DP

Parallelism layout of the main pretraining run as stated, as degree-annotated terms in the order TP, PP, VPP, EP, CP, DP ("PP16 · EP64 · ZeRO-1 DP"); a term without a stated degree has none ("FSDP").

2.664M H800 GPU-hours pretraining + 119K context extension (2.788M with 5K post-training)

Pretraining compute as stated (GPU-hours or FLOPs).

• main

  Tokens

  14800000000000

  Seq len

  4096

  Source: paper · p.22–23 §4.2, §4.3

  We set the maximum sequence length to 4K during pre-training, and pre-train DeepSeek-V3 on 14.8T tokens.

• context-1

  Seq len

  32768

  Lr

  7.3e-06

  Batch sequences

  1920

  Steps

  1000

  Source: paper · p.22–23 §4.2, §4.3

  In the first phase, the sequence length is set to 32K, and the batch size is 1920.

• context-2

  Seq len

  131072

  Lr

  7.3e-06

  Batch sequences

  480

  Steps

  1000

  Source: paper · p.22–23 §4.2, §4.3

  During the second phase, the sequence length is increased to 128K, and the batch size is reduced to 480.

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.

15.7B total (2.4B active) on 1.33T tokens; 228.7B total (20.9B active) on 540B tokens

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

22.0566

Calculated: train.tokens / params.total

train.tokens divided by params.total.

414.578

Calculated: train.tokens / params.counted_active_non_embedding

train.tokens divided by params.counted_active_non_embedding, or by the stated params.active_non_embedding when there is no validated checkpoint count.

0.193227

Calculated: train.batch_tokens / (train.lr_peak × train.weight_decay × train.tokens)

B / (η λ D) with B = train.batch_tokens, η = train.lr_peak, λ = train.weight_decay and D = train.tokens: the AdamW EMA timescale as a fraction of training (Bergsma et al. 2025, Power Lines, arXiv 2505.13738). Computed only when the optimizer is AdamW; uses the peak LR.

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.

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.

1,899 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).

248,937,610 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).

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.

8

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).