DeepSeek-V3.2

2025-12-01

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

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.

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.

671.9B parameters

Calculated: embedding + output head + all experts + other weights, deepseek-ai/DeepSeek-V3.2@a7e62ac 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.

1.853B parameters

Calculated: input embedding + output head, deepseek-ai/DeepSeek-V3.2@a7e62ac safetensors headers

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

670B parameters

Calculated: dense + routed backbone weights; excludes input embedding, output head, lookup tables, MTP and other modalities, deepseek-ai/DeepSeek-V3.2@a7e62ac 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.

11.52B parameters

Calculated: max(sum of dense and routed weights sharing a full layers.<index> tensor path), all experts included, deepseek-ai/DeepSeek-V3.2@a7e62ac 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.

36.55B parameters

Calculated: non-expert weights + routed experts × 8/256, excluding embedding and output head, deepseek-ai/DeepSeek-V3.2@a7e62ac 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.

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.

117.508

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).

61 DSA

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.

DSA: lightning indexer (64 heads × 128, ReLU, FP8), top-2048 KV tokens per query

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

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.

128

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

• MLA latent

  Layers

  61

  Elements

  512

  Source: code · inference/model.py; MLA.__init__ lines 541–542 and Indexer.__init__ lines 453–454

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

• RoPE key

  Layers

  61

  Elements

  64

  Source: code · inference/model.py; MLA.__init__ lines 541–542 and Indexer.__init__ lines 453–454

  self.register_buffer("pe_cache", torch.zeros(args.max_batch_size, args.max_seq_len, self.qk_rope_head_dim), persistent=False)

• DSA indexer key

  Layers

  61

  Elements

  128

  Source: code · inference/model.py; MLA.__init__ lines 541–542 and Indexer.__init__ lines 453–454

  self.register_buffer("k_cache", torch.zeros(args.max_batch_size, args.max_seq_len, self.head_dim, dtype=torch.float8_e4m3fn), persistent=False)

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.

• MLA latent

  Group

  MLA latent

  Elements

  512

  Bits

  8

  Format

  FP8 E4M3; FP32 scales per 128 channels

  Scale count

  4

  Scale bits

  32

  At

  README.md; V3.2 layout: 512 quantized latent bytes + 16 scale bytes

  Source: code · README.md; V3.2 layout: 512 quantized latent bytes + 16 scale bytes

  Next 16 bytes: Scale factors, containing 4 `float32` values.

  Serving layout: each layer retains 656 bytes of main MLA KV and 132 bytes of indexer K per token. 61 × (656 + 132) = 48,068 bytes/token, matching DeepSeek-V4.1 report p.1 Figure 1(b). The reference model simulates main-latent FP8 then retains BF16 buffers; this fact describes the packed FlashMLA serving layout.

• RoPE key

  Group

  RoPE key

  Elements

  64

  Bits

  16

  Format

  BF16

  At

  README.md; V3.2 unquantized RoPE key

  Source: code · README.md; V3.2 unquantized RoPE key

  Last 128 bytes: The "RoPE" part, containing 64 `bfloat16` values.

  Serving layout: each layer retains 656 bytes of main MLA KV and 132 bytes of indexer K per token. 61 × (656 + 132) = 48,068 bytes/token, matching DeepSeek-V4.1 report p.1 Figure 1(b). The reference model simulates main-latent FP8 then retains BF16 buffers; this fact describes the packed FlashMLA serving layout.

• DSA indexer key

  Group

  DSA indexer key

  Elements

  128

  Bits

  8

  Format

  FP8 E4M3; one FP32 scale

  Scale count

  1

  Scale bits

  32

  At

  inference/model.py lines 15, 453–454; block_size 128

  Source: code · inference/model.py lines 15, 453–454; block_size 128

  self.register_buffer("k_scale_cache", torch.zeros(args.max_batch_size, args.max_seq_len, self.head_dim // block_size, dtype=torch.float32), persistent=False)

  Serving layout: each layer retains 656 bytes of main MLA KV and 132 bytes of indexer K per token. 61 × (656 + 132) = 48,068 bytes/token, matching DeepSeek-V4.1 report p.1 Figure 1(b). The reference model simulates main-latent FP8 then retains BF16 buffers; this fact describes the packed FlashMLA serving layout.

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, ...).

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).

auxiliary-loss-free expert bias (added to scores for top-k selection only); coefficients not stated

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.

8 expert groups, top-4 groups per token (group score: sum of its top-2 biased expert scores)

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.

Not disclosed

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

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.

Not disclosed

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.

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.

lightning-indexer KL loss to the head-summed main attention distribution; indexer input detached from the LM loss

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

same distribution as the DeepSeek-V3.1-Terminus 128K long-context extension data

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

• dense warm-up (indexer only)

  Tokens

  2100000000

  Seq len

  131072

  Lr

  0.001

  Steps

  1000

  Batch sequences

  16

  Source: paper · p.4 §2.1.1

  We train the indexer for only 1000 steps, with each step consisting of 16 sequences of 128K tokens, resulting in a total of 2.1B tokens.

  Dense attention; all parameters frozen except the indexer, trained with a KL loss to the main attention distribution.

• sparse training

  Tokens

  943700000000

  Seq len

  131072

  Lr

  7.3e-06

  Steps

  15000

  Batch sequences

  480

  Source: paper · p.4 §2.1.1

  We train both the main model and the indexer for 15000 steps, with each step consisting of 480 sequences of 128K tokens, resulting in a total of 943.7B tokens.

  Top-2048 selection; all parameters trained; indexer input detached and trained only by the indexer KL loss.

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.

85,888 bytes

Calculated: kv.bytes_at_context / 131072 tokens; uniform BF16, payload only; quantization overhead excluded. MLA latent: 61 × (512 × 16/8) B/entry; RoPE key: 61 × (64 × 16/8) B/entry; DSA indexer key: 61 × (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.

85,888 bytes

Calculated: 62464 B / 1 tokens (MLA latent) + 7808 B / 1 tokens (RoPE key) + 15616 B / 1 tokens (DSA indexer key); uniform BF16, payload only; quantization overhead excluded. MLA latent: 61 × (512 × 16/8) B/entry; RoPE key: 61 × (64 × 16/8) B/entry; DSA indexer key: 61 × (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.

2,321 bytes

Calculated: kv.bytes_per_token (131072 tokens) / (params.active / 1e9); uniform BF16, payload only; quantization overhead excluded. MLA latent: 61 × (512 × 16/8) B/entry; RoPE key: 61 × (64 × 16/8) B/entry; DSA indexer key: 61 × (128 × 16/8) B/entry

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

304,257,079 bytes

Calculated: kv.bytes_at_context (131072 tokens) / (params.active / 1e9); uniform BF16, payload only; quantization overhead excluded. MLA latent: 61 × (512 × 16/8) B/entry; RoPE key: 61 × (64 × 16/8) B/entry; DSA indexer key: 61 × (128 × 16/8) B/entry

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

11,257,511,936 bytes

Calculated: 62464 B × 131072 (MLA latent) + 7808 B × 131072 (RoPE key) + 15616 B × 131072 (DSA indexer key); uniform BF16, payload only; quantization overhead excluded. MLA latent: 61 × (512 × 16/8) B/entry; RoPE key: 61 × (64 × 16/8) B/entry; DSA indexer key: 61 × (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.

21

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).