DeepSeek-V3.2

DeepSeek-AI · 2025-12-02 · analysed 2026-08-10 · skill v0.3.5

sparse-MoE efficient-attention text-only open-weights reasoning-specialised

Design overview

DeepSeek-V3.2 is a continued-trained member of the DeepSeek-V3 line: it starts from the DeepSeek-V3.1-Terminus base checkpoint (context already extended to 128K, architecture otherwise unchanged) and its only architectural modification is DeepSeek Sparse Attention (DSA). DSA replaces vanilla attention with a lightning indexer — a small number of heads computing a weighted ReLU score between each query and every preceding token, implementable in FP8 — and a fine-grained top-k selection that retrieves only the k=2048 best-scoring key-value entries per query, cutting the main attention cost from O(L²) to O(Lk). DSA is instantiated under MLA in its MQA mode (each latent KV vector shared across all query heads) for kernel efficiency; short-sequence prefilling uses a masked MHA mode that simulates DSA. The indexer is initialized in a frozen dense warm-up (1000 steps, 2.1B tokens) by KL-aligning its softmax scores to the summed head-wise attention distribution, then trained jointly in a 943.7B-token sparse stage. Post-training keeps the V3.2-Exp pipeline: six domain specialists plus writing and QA (thinking and non-thinking modes), distilled into the final model, then a single mixed GRPO stage fusing reasoning, agent, and alignment rewards, stabilized by an unbiased KL estimate, off-policy sequence masking, keep-routing, and keep-sampling masks; RL compute exceeds 10% of pre-training cost. Agentic capability comes from thinking context management (reasoning retained across tool messages) and a synthesis pipeline of 1,827 environments / 85,000 prompts. The relaxed-length DeepSeek-V3.2-Speciale variant, adding DeepSeekMath-V2 techniques, reaches gold-medal level at IMO/CMO/IOI 2025 and ranks 2nd at ICPC WF 2025.

What the report claims is novel

Atlas assessment: DSA is the first sparse-attention mechanism in the atlas built on MLA: an auxiliary lightning indexer computes cheap ReLU scores, a frozen dense warm-up (1000 steps) KL-aligns indexer softmax scores to the summed head-wise attention distribution, and top-k=2048 selection then sparsifies every layer's attention to O(Lk). It is a training-stabilized retrieval-style sparsification, not a new mixing family — contrast Kimi K3's KDA/MLA layerwise hybrid and the GLM-4.5 attention variants. The report's own parity evidence (base model matches V3.1-Terminus) is the key architectural claim and is internally consistent. The GRPO stability recipe (unbiased KL, off-policy masking, keep-routing, keep-sampling) is a substantive engineering contribution beyond deepseek-r1's report. Everything else — MLA, DeepSeekMoE, 128K context — is inherited unchanged from the V3 line (stated explicitly), so the atlas deepseek-v3 record remains the family's disclosure source for parameters. The >10%-of-pre-training RL compute and competition gold medals are internal claims, not independently verifiable.

Relation to the atlas

Lineage in

Influence out

Notable omissions

Closest relatives in the atlas

Spec sheet

Identity

FamilyDeepSeek-V3.2[1]
OrganisationDeepSeek-AI[2]
Release date2025-12-02[3]
Report URLhttps://arxiv.org/abs/2512.02556[4]
Licensenot disclosed
Open weightsyes

Scale

Total parametersnot disclosed
Active parameters / tokennot disclosed
Layersnot disclosed
Hidden dimensionnot disclosed
FFN inner dimensionnot disclosed
FFN rationot disclosed
Query headsnot disclosed
KV headsn/a[5]
Head dimensionnot disclosed
Vocabulary sizenot disclosed
Embedding tyingnot disclosed

Core block

Block typesparse-MoE[6]
Attention variantMLA[7]
Attention layer patternuniform
Depth mixingsequential-residual
Activationnot disclosed
Expert countnot disclosed
Experts per tokennot disclosed
Shared expertsnot disclosed
Routingnot disclosed
Load balancingnot disclosed
Expert granularitynot disclosed
Methodnot disclosed
RoPE base frequencynot disclosed
Partial RoPEnot disclosed
Typenot disclosed
Placementnot disclosed
QK-normnot disclosed
Attention sinksno
Softcappingno
Othernot disclosed

Context

Trained context length131,072[10]
Deployed context length131,072[9]
Extension methodnone

Tokenizer

Algorithmnot disclosed
Notesnot disclosed

Training

Training tokens945.8T
Data compositioncontinued pre-training data distribution 'totally aligned with the 128K long context extension data used for DeepSeek-V3.1-Terminus' (§2.1.1); post-training data: specialist-distilled domain data (mathematics, programming, general logical reasoning, general agentic tasks, agentic coding, agentic search, plus writing and general QA, in thinking and non-thinking modes) and agentic RL data (real tools: web search API, coding tools, Jupyter Notebooks; synthesized: 1,827 task-oriented environments with 4,417 general-agent tasks, plus 24,667 code-agent, 50,275 search-agent, 5,908 code-interpreter tasks)[11]
Curriculum / stagingtwo-stage DSA continued pre-training: (1) dense warm-up freezing all parameters except the lightning indexer, trained with a KL objective aligned to the summed head-wise attention distribution (LR 1e-3, 1000 steps); (2) sparse training with fine-grained top-k selection, optimizing all parameters (LR 7.3e-6, 15000 steps, k=2048); then post-training: specialist distillation → mixed GRPO RL (reasoning + agent + alignment in one stage) → continued RL to final checkpoints[12]
Optimizernot disclosed
LR schedulecontinued pre-training learning rates: 1e-3 (indexer warm-up stage), 7.3e-6 (sparse training stage)[10]
Batch schedule16 sequences of 128K tokens per step (dense warm-up); 480 sequences of 128K tokens per step (sparse stage)[10]
Precisionnot disclosed
Parallelismnot disclosed
Hardwarenot disclosed
Disclosed computenot disclosed

Post-training

SFTyes
Preference optimisationGRPO[13]
Reasoning trainingmixed RL merging reasoning, agent, and human-alignment training into one GRPO stage; rule-based outcome reward, length penalty and language consistency reward for reasoning/agent tasks, generative reward model with per-prompt rubrics for general tasks; RL compute >10% of pre-training cost; specialist distillation producing thinking (long-CoT) and non-thinking data; RL stabilizers: unbiased (importance-sampling corrected) KL estimate, off-policy sequence masking, keep routing, keep sampling mask; thinking-in-tool-use via cold-start (DeepSeek-V3 methodology) and thinking context management (reasoning retained across tool messages, discarded only on new user messages, tool-call history preserved); DeepSeek-V3.2-Speciale variant trained exclusively on reasoning data with reduced length penalty plus DeepSeekMath-V2 dataset/rewards for mathematical proofs[14]
Distillationspecialist distillation: per-domain specialist models (fine-tuned from the same pre-trained V3.2 base with large-scale RL) generate domain-specific data for the final checkpoint; 'models trained on the distilled data achieve performance levels only marginally below those of domain-specific specialists', with the gap eliminated by subsequent RL[13]

Modality

Typetext-only
Attachmentn/a[15]

Inference efficiency

KV-cache designMLA: low-rank latent KV cache; DSA retrieves only the top-k=2048 key-value entries per query token, so attention computation touches a sparse subset of the latent cache; MQA-mode MLA shares each latent vector across all query heads for kernel efficiency
Quantisation shippednone disclosed
Speculative decodingnone disclosed
Serving optimisationsmasked MHA mode simulating DSA for short-sequence prefilling (higher efficiency under short contexts), plus 'our optimized implementation' of DSA achieving significant end-to-end speedup in long-context scenarios; token-cost benchmarks from the deployed service on H800 clusters at $2 per GPU hour[16]

Evaluation

Benchmarks (report's own numbers only)

BenchmarkValueRef
MMLU-Pro (EM)85.0Table 2 (DeepSeek-V3.2 Thinking)
GPQA Diamond (Pass@1)82.4Table 2 (DeepSeek-V3.2 Thinking)
HLE (Pass@1)25.1Table 2 (DeepSeek-V3.2 Thinking)
LiveCodeBench (Pass@1-COT)83.3Table 2 (DeepSeek-V3.2 Thinking)
Codeforces (Rating)2386Table 2 (DeepSeek-V3.2 Thinking)
AIME 2025 (Pass@1)93.1Table 2 (DeepSeek-V3.2 Thinking)
HMMT Feb 2025 (Pass@1)92.5Table 2 (DeepSeek-V3.2 Thinking)
HMMT Nov 2025 (Pass@1)90.2Table 2 (DeepSeek-V3.2 Thinking)
IMOAnswerBench (Pass@1)78.3Table 2 (DeepSeek-V3.2 Thinking)
Terminal Bench 2.0 (Acc)46.4Table 2 (DeepSeek-V3.2 Thinking; Claude Code framework)
SWE Verified (Resolved)73.1Table 2 (DeepSeek-V3.2 Thinking)
SWE Multilingual (Resolved)70.2Table 2 (DeepSeek-V3.2 Thinking)
BrowseComp (Pass@1, with/without context management)51.4/67.6*Table 2 (DeepSeek-V3.2 Thinking)
BrowseCompZh (Pass@1)65.0Table 2 (DeepSeek-V3.2 Thinking)
HLE (Pass@1, search agent)40.8Table 2 (DeepSeek-V3.2 Thinking)
τ²-Bench (Pass@1)80.3Table 2 (DeepSeek-V3.2 Thinking)
MCP-Universe (Success Rate)45.9Table 2 (DeepSeek-V3.2 Thinking)
MCP-Mark (Pass@1)38.0Table 2 (DeepSeek-V3.2 Thinking)
Tool-Decathlon (Pass@1)35.2Table 2 (DeepSeek-V3.2 Thinking)
AIME 2025 (Pass@1, V3.2-Speciale)96.0Table 3 (DeepSeek-V3.2-Speciale)
GPQA Diamond (Pass@1, V3.2-Speciale)85.7Table 3 (DeepSeek-V3.2-Speciale)
HLE (Pass@1, V3.2-Speciale)30.6Table 3 (DeepSeek-V3.2-Speciale)
Codeforces (Rating, V3.2-Speciale)2701Table 3 (DeepSeek-V3.2-Speciale)
IMO 2025 (points, V3.2-Speciale)35/42 (Gold)Table 4 (DeepSeek-V3.2-Speciale)
IOI 2025 (points, V3.2-Speciale)492/600 (Gold, 10th place)Table 4 (DeepSeek-V3.2-Speciale)

Provenance references

  1. stated · Title
  2. stated · Title page byline (DeepSeek-AI)
  3. stated · arXiv header ('arXiv:2512.02556v1 [cs.CL] 02 Dec 2025'); confirmed via export.arxiv.org API (published 2025-12-02)
  4. stated · pipeline
  5. n/a · §2.1
  6. stated · §3.1
  7. stated · §2.1
  8. n/a · §2.1.1
  9. stated · §4.1
  10. stated · §2.1.1
  11. stated · §2.1.1, §3, Table 1
  12. stated · §2.1.1, §3
  13. stated · §3
  14. stated · §3, §3.1, §3.2
  15. n/a · §1
  16. stated · §2.3