DeepSeek-V4-Flash-Base

2026-04-24

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.

mit

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").

284B parameters

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

13B 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.

284.3B parameters

Calculated: embedding + output head + all experts + other weights, deepseek-ai/DeepSeek-V4-Flash-Base@8855555 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.

529.5M parameters

Calculated: input embedding + output head, deepseek-ai/DeepSeek-V4-Flash-Base@8855555 safetensors headers

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

283.8B parameters

Calculated: dense + routed backbone weights; excludes input embedding, output head, lookup tables, MTP and other modalities, deepseek-ai/DeepSeek-V4-Flash-Base@8855555 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.

6.596B parameters

Calculated: max(sum of dense and routed weights sharing a full layers.<index> tensor path), all experts included, deepseek-ai/DeepSeek-V4-Flash-Base@8855555 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.

13.27B parameters

Calculated: non-expert weights + routed experts × 6/256, excluding embedding and output head, deepseek-ai/DeepSeek-V4-Flash-Base@8855555 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.

43

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

4,096

Residual stream width.

129,280

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

No

Input embedding and output head share weights.

DeepSeek-V3 byte-level BPE plus special tokens for context construction; 128K vocab

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

95.2558

Calculated: d_model / layers

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

SWA, CSA, HCA (CSA and HCA each with a 128-token SWA branch)

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

2 SWA, then 1 CSA : 1 HCA (21 CSA + 20 HCA)

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.

41

Inferred from the source; see the explanation below.

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

0.0487805

Calculated: attn.local_layers / attn.global_layers

attn.local_layers divided by attn.global_layers, the sliding-window to global interleave ratio (5 for '5 SWA : 1 global').

2

Layers with a fixed attention window.

64

Query heads in the global-attention layers.

1

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

32,768

Calculated: heads × head_dim

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

64

Calculated: heads / kv_heads

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

512

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

128 tokens

Sliding-window size of the local layers.

1

KV heads in sliding-window layers (stated even when equal to attn.kv_heads).

64

Query heads in sliding-window layers (stated even when equal to attn.heads).

512

Query/key head dimension in sliding-window layers. Value head dimensions may differ; see attn.local_v_head_dim.

1,024

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

partial RoPE (last 64 dims of Q and KV entries; inverse rotation on attention output)

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

160,000

RoPE base frequency of the global layers at release.

10,000

RoPE base frequency of the sliding-window layers (stated even when equal to attn.rope_theta).

0.125

Inferred from the source; see the explanation below.

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

Yes

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

CSA: lightning indexer (64 heads × 128) selects top-512 of 4×-compressed KV entries; HCA (128×) attends densely

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

learnable per-head sink logit in the softmax denominator, every layer (CSA, HCA and the 2 pure-SWA layers)

Learned sink logits or sink bias in the softmax, with the layer types that use them.

No

Query/key/value projections have bias terms.

• CSA main KV

  Layers

  21

  Elements

  512

  Tokens per entry

  4

  Source: code · V4-Flash inference/model.py; Attention.__init__, Attention.forward, Indexer.__init__

  kv_cache_size = args.window_size + (args.max_seq_len // self.compress_ratio if self.compress_ratio else 0)

  Layers with /compress_ratios 4; one shared K=V entry of head_dim per 4 tokens.

• CSA indexer K

  Layers

  21

  Elements

  128

  Tokens per entry

  4

  Source: code · V4-Flash inference/model.py; Attention.__init__, Attention.forward, Indexer.__init__

  self.register_buffer("kv_cache", torch.zeros(args.max_batch_size, args.max_seq_len // compress_ratio, self.head_dim), persistent=False)

  One index_head_dim key per compressed entry (Indexer.kv_cache).

• HCA main KV

  Layers

  20

  Elements

  512

  Tokens per entry

  128

  Source: code · V4-Flash inference/model.py; Attention.__init__, Attention.forward, Indexer.__init__

  self.register_buffer("kv_cache", torch.zeros(args.max_batch_size, kv_cache_size, self.head_dim), persistent=False)

  Layers with /compress_ratios 128; no indexer.

• SWA window

  Layers

  43

  Elements

  512

  Window

  128

  Source: code · V4-Flash inference/model.py; Attention.__init__, Attention.forward, Indexer.__init__

  self.kv_cache[:bsz, start_pos % win] = kv.squeeze(1)

  Ring buffer in every backbone layer (the 2 pure-SWA layers and the SWA branch of CSA/HCA); MTP block excluded.

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.

• CSA main KV NoPE

  Group

  CSA main KV

  Elements

  448

  Bits

  8

  Format

  FP8 E4M3; E8M0 scales per 64 channels

  Scale count

  7

  Scale bits

  8

  Padding bytes

  1

  At

  README.md; V4 layout: 448 NoPE bytes, 7 scale bytes and 1 padding byte

  Source: code · README.md; V4 layout: 448 NoPE bytes, 7 scale bytes and 1 padding byte

  The scale row is 8 Bytes

  Serving layout from the official FlashMLA kernels and post-training FP4 indexer QAT; main KV keeps the RoPE tail in BF16. Each main/SWA entry is 584 bytes, each indexer entry 68 bytes. This serving recipe is documented for the post-trained V4-Flash sibling; it does not establish FP4 QAT of the base checkpoint.

• CSA main KV RoPE

  Group

  CSA main KV

  Elements

  64

  Bits

  16

  Format

  BF16

  At

  README.md; V4 layout: 128 RoPE bytes

  Source: code · README.md; V4 layout: 128 RoPE bytes

  the 64 `bfloat16` RoPE values (not quantized)

  Serving layout from the official FlashMLA kernels and post-training FP4 indexer QAT; main KV keeps the RoPE tail in BF16. Each main/SWA entry is 584 bytes, each indexer entry 68 bytes. This serving recipe is documented for the post-trained V4-Flash sibling; it does not establish FP4 QAT of the base checkpoint.

• HCA main KV NoPE

  Group

  HCA main KV

  Elements

  448

  Bits

  8

  Format

  FP8 E4M3; E8M0 scales per 64 channels

  Scale count

  7

  Scale bits

  8

  Padding bytes

  1

  At

  README.md; V4 layout: 448 NoPE bytes, 7 scale bytes and 1 padding byte

  Source: code · README.md; V4 layout: 448 NoPE bytes, 7 scale bytes and 1 padding byte

  The scale row is 8 Bytes

  Serving layout from the official FlashMLA kernels and post-training FP4 indexer QAT; main KV keeps the RoPE tail in BF16. Each main/SWA entry is 584 bytes, each indexer entry 68 bytes. This serving recipe is documented for the post-trained V4-Flash sibling; it does not establish FP4 QAT of the base checkpoint.

• HCA main KV RoPE

  Group

  HCA main KV

  Elements

  64

  Bits

  16

  Format

  BF16

  At

  README.md; V4 layout: 128 RoPE bytes

  Source: code · README.md; V4 layout: 128 RoPE bytes

  the 64 `bfloat16` RoPE values (not quantized)

  Serving layout from the official FlashMLA kernels and post-training FP4 indexer QAT; main KV keeps the RoPE tail in BF16. Each main/SWA entry is 584 bytes, each indexer entry 68 bytes. This serving recipe is documented for the post-trained V4-Flash sibling; it does not establish FP4 QAT of the base checkpoint.

• SWA window NoPE

  Group

  SWA window

  Elements

  448

  Bits

  8

  Format

  FP8 E4M3; E8M0 scales per 64 channels

  Scale count

  7

  Scale bits

  8

  Padding bytes

  1

  At

  README.md; V4 layout: 448 NoPE bytes, 7 scale bytes and 1 padding byte

  Source: code · README.md; V4 layout: 448 NoPE bytes, 7 scale bytes and 1 padding byte

  The scale row is 8 Bytes

  Serving layout from the official FlashMLA kernels and post-training FP4 indexer QAT; main KV keeps the RoPE tail in BF16. Each main/SWA entry is 584 bytes, each indexer entry 68 bytes. This serving recipe is documented for the post-trained V4-Flash sibling; it does not establish FP4 QAT of the base checkpoint.

• SWA window RoPE

  Group

  SWA window

  Elements

  64

  Bits

  16

  Format

  BF16

  At

  README.md; V4 layout: 128 RoPE bytes

  Source: code · README.md; V4 layout: 128 RoPE bytes

  the 64 `bfloat16` RoPE values (not quantized)

  Serving layout from the official FlashMLA kernels and post-training FP4 indexer QAT; main KV keeps the RoPE tail in BF16. Each main/SWA entry is 584 bytes, each indexer entry 68 bytes. This serving recipe is documented for the post-trained V4-Flash sibling; it does not establish FP4 QAT of the base checkpoint.

• CSA indexer key

  Group

  CSA indexer K

  Elements

  128

  Bits

  4

  Format

  MXFP4 E2M1; E8M0 scales per 32 channels

  Scale count

  4

  Scale bits

  8

  At

  inference/kernel.py; fp4_act_quant, default block_size 32

  Source: code · inference/kernel.py; fp4_act_quant, default block_size 32

  s = z.new_empty(*z.size()[:-1], N // block_size, dtype=torch.float8_e8m0fnu)

  Serving layout from the official FlashMLA kernels and post-training FP4 indexer QAT; main KV keeps the RoPE tail in BF16. Each main/SWA entry is 584 bytes, each indexer entry 68 bytes. This serving recipe is documented for the post-trained V4-Flash sibling; it does not establish FP4 QAT of the base checkpoint.

Sourced storage components of attn.kv_groups. Each component names its group and gives elements, bits per element, format, scale_count, scale_bits and padding_bytes per entry. Different parts of one entry can have different formats. Sourced-storage sizes are available only when every group's elements are covered. These are inference-cache formats; they do not establish pretraining precision.

SwiGLU

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

up (linear) clamped to [−10, 10], gate capped at 10; routed and shared experts in every layer

Clamping of GLU gate / up activations (e.g. SwiGLU limit 7), with the layers it applies to.

No

FFN projections have bias terms.

256

Routed experts per MoE layer.

6

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).

0

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

sqrt(softplus); hash on token ID (first 3 MoE layers)

Gating function over expert logits (softmax, sigmoid).

aux-loss-free bias (update speed 0.001) + sequence-wise loss 1e-4; no node-limited routing

Load-balancing method, with coefficients where disclosed.

1.5

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

Yes

Selected expert weights are renormalized to sum to 1.

0.5

Calculated: moe.expert_dim / d_model

moe.expert_dim divided by shape.d_model.

42.6667

Calculated: experts / top_k

Routed experts divided by experts selected per token.

0.0457746

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.

e_proj(RMSNorm(Emb(t_{i+1}))) + h_proj(RMSNorm(h)), then one block with 128-token pure-SWA attention and MoE FFN inside mHC, own mHC head and final norm

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.

0.3, reduced to 0.1 at the start of LR decay

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.

mHC: 4 residual streams, residual mixing projected to doubly stochastic by Sinkhorn-Knopp (20 iterations), sigmoid-bounded input/output maps

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

4.096K tokens

Sequence length of the main pretraining stage.

1.049M tokens

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.

4K → 16K → 64K → 1M during pretraining; compressed layers use YaRN (factor 16, original 65536)

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

32T tokens

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

Muon + AdamW (AdamW for embedding, prediction head and RMSNorm weights)

Optimizer used for the bulk of parameters.

0.00027

Peak learning rate of the main pretraining stage.

2.7e-05

Learning rate at the end of the decay.

linear warmup (2000 steps), constant 2.7e-4 for most of training, cosine decay to 2.7e-5 near the end

Learning-rate schedule shape with phase boundaries.

2000 steps

Warmup length in steps or tokens.

ramped from a small size to 75.5M tokens, then constant for most of training

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

75.5M 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.

Anticipatory Routing (routing indices computed with parameters from Δt steps earlier), auto-triggered with a short rollback on loss spikes; SwiGLU clamping

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

sample-level attention masking within packed sequences

Whether attention crosses document boundaries within a packed sequence.

FIM, strategy inherited from DeepSeek-V3

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

>32T tokens: math, code, web, long documents (scientific papers, technical reports), larger multilingual corpus; agentic data added in mid-training

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

• context-4096

  Seq len

  4096

  Source: paper · p.25 §4.2.2

  The training starts with a sequence length of 4K

  Dense attention.

• context-16384

  Seq len

  16384

  Source: paper · p.25 §4.2.2

  we gradually extend the training sequence length to 16K, 64K, and 1M.

• context-65536

  Seq len

  65536

  Source: paper · p.25 §4.2.2

  introduce sparse attention at the sequence length of 64K and keep sparse attention during the rest of the training.

  Sparse attention introduced here after a short indexer warm-up and dense attention for the first 1T tokens; kept for the rest of training.

• context-1048576

  Seq len

  1048576

  Source: paper · p.25 §4.2.2

  we gradually extend the training sequence length to 16K, 64K, and 1M.

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.

AdamW hyper-parameters reused for Muon by rescaling the update RMS

How peak LR, batch size or other hyperparameters were chosen for the released run (fitted scaling law with its formula, μP transfer from a proxy width, sweep), as stated.

112.676

Calculated: train.tokens / params.total

train.tokens divided by params.total.

2,411.01

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.

6,923 bytes

Calculated: kv.bytes_at_context / 131072 tokens; uniform BF16, payload only; quantization overhead excluded. CSA main KV: 21 × (512 × 16/8) B/entry; CSA indexer K: 21 × (128 × 16/8) B/entry; HCA main KV: 20 × (512 × 16/8) B/entry; SWA window: 43 × (512 × 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.

6,880 bytes

Calculated: 21504 B / 4 tokens (CSA main KV) + 5376 B / 4 tokens (CSA indexer K) + 20480 B / 128 tokens (HCA main KV); uniform BF16, payload only; quantization overhead excluded. CSA main KV: 21 × (512 × 16/8) B/entry; CSA indexer K: 21 × (128 × 16/8) B/entry; HCA main KV: 20 × (512 × 16/8) B/entry; SWA window: 43 × (512 × 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.

533 bytes

Calculated: kv.bytes_per_token (131072 tokens) / (params.active / 1e9); uniform BF16, payload only; quantization overhead excluded. CSA main KV: 21 × (512 × 16/8) B/entry; CSA indexer K: 21 × (128 × 16/8) B/entry; HCA main KV: 20 × (512 × 16/8) B/entry; SWA window: 43 × (512 × 16/8) B/entry

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

69,800,881 bytes

Calculated: kv.bytes_at_context (131072 tokens) / (params.active / 1e9); uniform BF16, payload only; quantization overhead excluded. CSA main KV: 21 × (512 × 16/8) B/entry; CSA indexer K: 21 × (128 × 16/8) B/entry; HCA main KV: 20 × (512 × 16/8) B/entry; SWA window: 43 × (512 × 16/8) B/entry

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

907,411,456 bytes

Calculated: 21504 B × floor(131072 / 4) (CSA main KV) + 5376 B × floor(131072 / 4) (CSA indexer K) + 20480 B × floor(131072 / 128) (HCA main KV) + 44032 B × min(131072, 128) (SWA window); uniform BF16, payload only; quantization overhead excluded. CSA main KV: 21 × (512 × 16/8) B/entry; CSA indexer K: 21 × (128 × 16/8) B/entry; HCA main KV: 20 × (512 × 16/8) B/entry; SWA window: 43 × (512 × 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.

24

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).