DeepSeek-V4.1-Flash

2026-09-10

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 + 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").

552B parameters

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

16B parameters

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

8B parameters

Parameters used per prompt token during prefill when the publisher states it differs from decode (e.g. YOCO-style designs that skip the upper half of the stack).

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.

551.6B parameters

Calculated: embedding + output head + all experts + other weights, deepseek-ai/DeepSeek-V4.1-Flash@314113f safetensors headers; reproduces stated total (other modalities) 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.

661.9M parameters

Calculated: input embedding + output head, deepseek-ai/DeepSeek-V4.1-Flash@314113f safetensors headers

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

196.9B parameters

Calculated: n-gram / per-layer embedding tables, deepseek-ai/DeepSeek-V4.1-Flash@314113f safetensors headers

Parameters in n-gram / Engram / per-layer embedding tables counted from the checkpoint. Like the token embedding, they are lookups and are excluded from params.counted_total and params.counted_active_non_embedding.

551B parameters

Calculated: dense + routed backbone weights; excludes input embedding, output head, lookup tables, MTP and other modalities, deepseek-ai/DeepSeek-V4.1-Flash@314113f safetensors headers; reproduces stated total (other modalities) 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.

13.77B parameters

Calculated: max(sum of dense and routed weights sharing a full layers.<index> tensor path), all experts included, deepseek-ai/DeepSeek-V4.1-Flash@314113f 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.

15.89B parameters

Calculated: non-expert weights + routed experts × 6/384, excluding embedding and output head, deepseek-ai/DeepSeek-V4.1-Flash@314113f safetensors headers; reproduces stated total (other modalities) 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.

Backbone only; the 196B Engram (conditional-memory) parameters are reported separately.

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

40

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

5,120

Residual stream width.

129,280

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

No

Input embedding and output head share weights.

Engram, 196B params in 2 modules at layers 1 and 14 (0-indexed); N-gram orders {2, 3, 4}, 8 hash heads, 2048 dims per order, ~16M-entry prime-sized tables, FP8, no short conv

Hashed n-gram or conditional-memory tables outside the dense backbone (Engram, n-gram embeddings), with size and placement.

128

Calculated: d_model / layers

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

SWA, CSA2

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

2 SWA + 18 CSA2 m=2 (encoder: 3 × [1 Full, 5 Reuse]) + 20 CSA2 m=1 (decoder: [1 Full, 3 Reuse], then 4 × [1 Reindex, 3 Reuse])

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.

38

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.

0.0526316

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.

1,280

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; 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.

No

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

CSA2 indexer (32 heads × 128), top-512 KV entries; decoder hierarchical indexer limited to a 2048-block × 8-position candidate pool

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

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

No

Query/key/value projections have bias terms.

Main KV and indexer K owned by layers 2, 8, 14 (encoder, m=2, each serving its 6-layer group) and 20 (decoder, m=1, serving layers 20–39, computed from the last encoder layer's output); other CSA2 layers read them. Each layer keeps its own 128-token SWA KV.

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

• encoder main KV (m=2)

  Layers

  3

  Elements

  512

  Tokens per entry

  2

  Source: code · inference/model.py; Attention.__init__, Compressor.forward, Indexer.__init__

  compress_ratio > 0 does not mean the layer compresses its own KV: only kv_source_layers do, the rest read that same cache.

  Layers 2, 8, 14; one shared K=V entry per 2 tokens.

• encoder indexer K (m=2)

  Layers

  3

  Elements

  128

  Tokens per entry

  2

  Source: code · inference/model.py; Attention.__init__, Compressor.forward, Indexer.__init__

  torch.zeros(args.max_batch_size, args.max_seq_len // self.compress_ratio, args.index_head_dim)

  Indexer.k_cache of the same owner layers.

• decoder main KV (m=1)

  Layers

  1

  Elements

  512

  Tokens per entry

  1

  Source: code · inference/model.py; Attention.__init__, Compressor.forward, Indexer.__init__

  if ratio == 1: # one token per group: nothing to pool, so no gate and no fp32

  Layer 20; ratio-1 compressor is a plain projection.

• decoder indexer K (m=1)

  Layers

  1

  Elements

  128

  Tokens per entry

  1

  Source: code · inference/model.py; Attention.__init__, Compressor.forward, Indexer.__init__

  self.owns_k = layer_id in args.kv_source_layers

  Layer 20.

• SWA window

  Layers

  40

  Elements

  512

  Window

  128

  Source: code · inference/model.py; Attention.__init__, Compressor.forward, Indexer.__init__

  torch.zeros(args.max_batch_size, args.window_size, self.head_dim)

  window_kv_cache in every backbone layer; DSpark layers 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.

• encoder main KV (m=2)

  Group

  encoder main KV (m=2)

  Elements

  512

  Bits

  4

  Format

  FP4 E2M1; E4M3 scales per 16 channels

  Scale count

  32

  Scale bits

  8

  At

  README.md; V4.1 fp4 layout: 256 data bytes + 32 scale bytes

  Source: code · README.md; V4.1 fp4 layout: 256 data bytes + 32 scale bytes

  The scale row is 32 Bytes of `float8_e4m3`

  Post-training/serving recipe, not pretraining precision. Main KV: 288 bytes/entry; indexer K: 68; SWA: 528. Global growth = (3/2 + 1) × (288 + 68) = 890 bytes/token, matching report p.1 Figure 1(b). The finite SWA rings are additional runtime cache, not part of that global figure. FP4 main-cache QAT is introduced in post-training (§2.4.4).

• decoder main KV (m=1)

  Group

  decoder main KV (m=1)

  Elements

  512

  Bits

  4

  Format

  FP4 E2M1; E4M3 scales per 16 channels

  Scale count

  32

  Scale bits

  8

  At

  README.md; V4.1 fp4 layout: 256 data bytes + 32 scale bytes

  Source: code · README.md; V4.1 fp4 layout: 256 data bytes + 32 scale bytes

  The scale row is 32 Bytes of `float8_e4m3`

  Post-training/serving recipe, not pretraining precision. Main KV: 288 bytes/entry; indexer K: 68; SWA: 528. Global growth = (3/2 + 1) × (288 + 68) = 890 bytes/token, matching report p.1 Figure 1(b). The finite SWA rings are additional runtime cache, not part of that global figure. FP4 main-cache QAT is introduced in post-training (§2.4.4).

• encoder indexer K (m=2)

  Group

  encoder indexer K (m=2)

  Elements

  128

  Bits

  4

  Format

  MXFP4 E2M1; E8M0 scales per 32 channels

  Scale count

  4

  Scale bits

  8

  At

  inference/model.py lines 28–30, 546; fp4_act_quant(k, fp4_block_size, True)

  Source: code · inference/model.py lines 28–30, 546; fp4_act_quant(k, fp4_block_size, True)

  fp4_block_size = 32

  Post-training/serving recipe, not pretraining precision. Main KV: 288 bytes/entry; indexer K: 68; SWA: 528. Global growth = (3/2 + 1) × (288 + 68) = 890 bytes/token, matching report p.1 Figure 1(b). The finite SWA rings are additional runtime cache, not part of that global figure. FP4 main-cache QAT is introduced in post-training (§2.4.4).

• decoder indexer K (m=1)

  Group

  decoder indexer K (m=1)

  Elements

  128

  Bits

  4

  Format

  MXFP4 E2M1; E8M0 scales per 32 channels

  Scale count

  4

  Scale bits

  8

  At

  inference/model.py lines 28–30, 546; fp4_act_quant(k, fp4_block_size, True)

  Source: code · inference/model.py lines 28–30, 546; fp4_act_quant(k, fp4_block_size, True)

  fp4_block_size = 32

  Post-training/serving recipe, not pretraining precision. Main KV: 288 bytes/entry; indexer K: 68; SWA: 528. Global growth = (3/2 + 1) × (288 + 68) = 890 bytes/token, matching report p.1 Figure 1(b). The finite SWA rings are additional runtime cache, not part of that global figure. FP4 main-cache QAT is introduced in post-training (§2.4.4).

• SWA window

  Group

  SWA window

  Elements

  512

  Bits

  8

  Format

  FP8 E4M3; E8M0 scales per 32 channels

  Scale count

  16

  Scale bits

  8

  At

  README.md; V4.1 FP8 layout: 512 data bytes + 16 scale bytes

  Source: code · README.md; V4.1 FP8 layout: 512 data bytes + 16 scale bytes

  The scale row is 16 Bytes of `float8_e8m0`

  Post-training/serving recipe, not pretraining precision. Main KV: 288 bytes/entry; indexer K: 68; SWA: 528. Global growth = (3/2 + 1) × (288 + 68) = 890 bytes/token, matching report p.1 Figure 1(b). The finite SWA rings are additional runtime cache, not part of that global figure. FP4 main-cache QAT is introduced in post-training (§2.4.4).

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.

384

Routed experts per MoE layer.

6

Routed experts selected per token.

1

Always-on experts per MoE layer.

2,304

Intermediate width of one routed expert.

2,304

Total intermediate width of the shared expert(s).

0

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

sqrt(softplus)

Gating function over expert logits (softmax, sigmoid).

aux-loss-free bias with separate text and image biases (update speed 0.001) + sequence-level loss 1e-4

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.

0.45

Calculated: moe.expert_dim / d_model

moe.expert_dim divided by shape.d_model.

64

Calculated: experts / top_k

Routed experts divided by experts selected per token.

0.0289855

Calculated: params.active / params.total

params.active divided by params.total.

0

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).

1e-20

Epsilon of the block normalization layers.

Single-Pass mHC: 4 residual streams, Sinkhorn-Knopp (20 iterations); each sublayer's input mixing uses the coefficients predicted at the previous sublayer

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

65.54K 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.

64K → 1M at 34T tokens; CSA2 layers use YaRN (factor 16, original 65536)

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

45T tokens

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

Muon + AdamW (Muon for linear weights, head-wise for Q/K; AdamW for RMSNorm and other non-matrix params; Sinkhorn-balanced momentum update for embeddings, Engram tables and prediction head)

Optimizer used for the bulk of parameters.

0.00026

Peak learning rate of the main pretraining stage.

2.6e-05

Learning rate at the end of the decay.

2000-step linear warmup, constant 2.6e-4 to 28T, cosine to 2.6e-5 at 40T, constant to 45T

Learning-rate schedule shape with phase boundaries.

2000 steps

Warmup length in steps or tokens.

100.6M tokens, fixed

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

100.6M 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.

sample-level attention masking within packed sequences

Whether attention crosses document boundaries within a packed sequence.

multimodal (image + text) input from the start of pretraining

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

text-only : multimodal = 7:1 tokens; multimodal = image-text pairs, interleaved image-text, domain-specific (grounding, OCR, image-code, computer-use)

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

• main

  Tokens

  34000000000000

  Seq len

  65536

  Source: paper · p.22 §4.2.2

  We train the model from scratch with sparse attention at a sequence length of 64K

  Sparse attention from scratch.

• long

  Tokens

  11000000000000

  Seq len

  1048576

  Source: paper · p.22 §4.2.2

  extend the sequence length to 1M at 34T tokens.

  34T to 45T.

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 LR reused for Muon by rescaling update RMS to 0.18

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.

ladder over model parameters and training data (sizes not disclosed)

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

81.5217

Calculated: train.tokens / params.total

train.tokens divided by params.total.

2,832.08

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.

3,240 bytes

Calculated: kv.bytes_at_context / 131072 tokens; uniform BF16, payload only; quantization overhead excluded. encoder main KV (m=2): 3 × (512 × 16/8) B/entry; encoder indexer K (m=2): 3 × (128 × 16/8) B/entry; decoder main KV (m=1): 1 × (512 × 16/8) B/entry; decoder indexer K (m=1): 1 × (128 × 16/8) B/entry; SWA window: 40 × (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.

3,200 bytes

Calculated: 3072 B / 2 tokens (encoder main KV (m=2)) + 768 B / 2 tokens (encoder indexer K (m=2)) + 1024 B / 1 tokens (decoder main KV (m=1)) + 256 B / 1 tokens (decoder indexer K (m=1)); uniform BF16, payload only; quantization overhead excluded. encoder main KV (m=2): 3 × (512 × 16/8) B/entry; encoder indexer K (m=2): 3 × (128 × 16/8) B/entry; decoder main KV (m=1): 1 × (512 × 16/8) B/entry; decoder indexer K (m=1): 1 × (128 × 16/8) B/entry; SWA window: 40 × (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.

202 bytes

Calculated: kv.bytes_per_token (131072 tokens) / (params.active / 1e9); uniform BF16, payload only; quantization overhead excluded. encoder main KV (m=2): 3 × (512 × 16/8) B/entry; encoder indexer K (m=2): 3 × (128 × 16/8) B/entry; decoder main KV (m=1): 1 × (512 × 16/8) B/entry; decoder indexer K (m=1): 1 × (128 × 16/8) B/entry; SWA window: 40 × (512 × 16/8) B/entry

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

26,542,080 bytes

Calculated: kv.bytes_at_context (131072 tokens) / (params.active / 1e9); uniform BF16, payload only; quantization overhead excluded. encoder main KV (m=2): 3 × (512 × 16/8) B/entry; encoder indexer K (m=2): 3 × (128 × 16/8) B/entry; decoder main KV (m=1): 1 × (512 × 16/8) B/entry; decoder indexer K (m=1): 1 × (128 × 16/8) B/entry; SWA window: 40 × (512 × 16/8) B/entry

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

424,673,280 bytes

Calculated: 3072 B × floor(131072 / 2) (encoder main KV (m=2)) + 768 B × floor(131072 / 2) (encoder indexer K (m=2)) + 1024 B × 131072 (decoder main KV (m=1)) + 256 B × 131072 (decoder indexer K (m=1)) + 40960 B × min(131072, 128) (SWA window); uniform BF16, payload only; quantization overhead excluded. encoder main KV (m=2): 3 × (512 × 16/8) B/entry; encoder indexer K (m=2): 3 × (128 × 16/8) B/entry; decoder main KV (m=1): 1 × (512 × 16/8) B/entry; decoder indexer K (m=1): 1 × (128 × 16/8) B/entry; SWA window: 40 × (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.

39

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).