Production model

DeepSeek V3.2-Exp

DeepSeek-AI · Released 2025-09-29 · 671B parameters (37B active) · 128,000 ctx · open weights

Architecture

Positional encoding	RoPE (decoupled head for MLA)
Norm placement	Pre-Norm
Norm type	RMSNorm
QK-Norm	no
Activation	SwiGLU
Attention	MLA + DeepSeek Sparse Attention (Lightning Indexer + top-K key selection)
MoE	DeepSeekMoE with aux-loss-free routing
Other	Same V3 base: 671B total, 37B active, 61 layers, First production model with DSA — sparse attention via learned top-K selection, 'Exp' label: experimental release, not the mainline V4

Techniques used

DeepSeek Sparse Attention — First DSA-enabled DeepSeek release; experimental V3 derivative that ships sparse attention as the long-context efficiency lever.
Multi-Head Latent Attention — MLA with the DeepSeek Sparse Attention layer added on top (Lightning Indexer + top-K selection).
Auxiliary-Loss-Free Load Balancing — Aux-loss-free routing carried over from V3 / V3.1; only attention layers change in V3.2.

DeepSeek V3.2-Exp (September 2025) is the experimental V3 derivative that ships DeepSeek Sparse Attention (DSA) as the long-context efficiency mechanism. Architecturally it is V3 plus one new layer: the Lightning Indexer + top-K key selection that sits on top of MLA.

Everything else carries over from DeepSeek V3:

671B total parameters, 37B active per token
61 transformer layers
MLA attention with a decoupled RoPE head
DeepSeekMoE with aux-loss-free routing
Pre-Norm RMSNorm, SwiGLU FFN
128K context window via YaRN extension

The “Exp” label signals that DeepSeek treats this as an experimental release rather than the mainline V4. The DSA mechanism is novel enough that long-term behavior at extreme contexts and on retrieval-heavy tasks is still under study; V3.2 lets the open community evaluate the approach at frontier scale before it gets baked into a fully numbered release.

For the architectural details specific to V3.2 — how the Lightning Indexer is trained, how top-K selection interacts with the MLA latent cache, the cost/quality tradeoff at 128K — see the DSA technique entry.

Sources

Export

BibTeX

@article{deepseekai2025,
  title         = {DeepSeek V3.2-Exp},
  author        = {DeepSeek-AI},
  year          = {2025},
  url           = {https://github.com/deepseek-ai/DeepSeek-V3.2-Exp/blob/main/DeepSeek_V3_2.pdf}
}

CSL JSON

{
  "id": "deepseekai2025",
  "type": "article-journal",
  "title": "DeepSeek V3.2-Exp",
  "author": [
    {
      "literal": "DeepSeek-AI"
    }
  ],
  "issued": {
    "date-parts": [
      [
        2025
      ]
    ]
  },
  "URL": "https://github.com/deepseek-ai/DeepSeek-V3.2-Exp/blob/main/DeepSeek_V3_2.pdf"
}

RIS

TY  - JOUR
TI  - DeepSeek V3.2-Exp
AU  - DeepSeek-AI
PY  - 2025
UR  - https://github.com/deepseek-ai/DeepSeek-V3.2-Exp/blob/main/DeepSeek_V3_2.pdf
ER  -

Models like this