Paper Breakdown: DeepSeek-V4

Summary: DeepSeek-V4 introduces hybrid CSA+HCA attention for 1M token contexts, Manifold-Constrained Hyper-Connections (mHC) for stable deep residual connections, and Muon optimizer with hybrid Newton-Schulz for faster convergence. V4-Pro (1.6T params, 49B activated) and V4-Flash (284B params, 13B activated) use FP4 experts and achieve 27% FLOPs / 10% KV cache vs V3.2 at 1M tokens. Post-training uses independent expert cultivation + on-policy distillation.


1. Key Contributions

# Contribution Section Impact
1 Hybrid CSA + HCA Attention 2.3 1M context; 27% FLOPs, 10% KV cache vs V3.2
2 mHC (Manifold Hyper-Connections) 2.2 Stable deep residual connections via Birkhoff polytope
3 Muon Optimizer (Hybrid Newton-Schulz) 2.4 Faster convergence, better stability for matrix params
4 Infrastructure: Fused MoE kernel, TileLang, FP4, heterogeneous KV cache 3 Training/inference efficiency
5 Post-training: Expert cultivation + on-policy distillation 3.6 Unified model from domain specialists

2. Architecture Overview (Fig 2)

DeepSeek-V4 Block:
┌─────────────────────────────┐
│  mHC (Manifold HC)          │  ← Input residual
├─────────────────────────────┤
│  Attention (CSA or HCA)     │  ← Alternating layers
├─────────────────────────────┤
│  mHC                        │  ← Post-attention residual
├─────────────────────────────┤
│  DeepSeekMoE FFN            │  ← Mixture of Experts
├─────────────────────────────┤
│  mHC                        │  ← Post-FFN residual
└─────────────────────────────┘
  • Retains: DeepSeekMoE, MTP (Multi-Token Prediction)
  • New: CSA/HCA hybrid, mHC, Muon

3. mHC: Manifold-Constrained Hyper-Connections (§2.2)

Standard HC Problem

  • Expand residual: R^d to R^{n_{hc} × d}
  • X_{l+1} = B_l X_l + C_l F_l(A_l X_l)
  • B_l unconstrained → spectral norm unbounded → instability in deep stacks

mHC Solution

Constrain B_l to Birkhoff polytope M (doubly stochastic matrices):

B_l ∈ M = \{M mid Mmathbf{1}=mathbf{1},\; mathbf{1}^T M = mathbf{1}^T,\; M ≥ 0\}

||B_l||_2 ≤ 1 (non-expansive) → stable propagation

A_l, C_l constrained via Sigmoid: non-negativity + boundedness

Dynamic Parameterization:

  • Input X_l to hat{X}_l = RMSNorm(vec(X_l))
  • tilde{A}_l = alpha_l^pre(hat{X}_l W^pre_l) + S^pre_l
  • tilde{B}_l = alpha_l^resMat(hat{X}_l W^res_l) + S^res_lSinkhorn-KnoppB_l
  • tilde{C}_l = alpha_l^post(hat{X}_l W^post_l)^T + S^post_l → Sigmoid → C_l

Sinkhorn-Knopp (20 iters):

M^{(0)} = exp(tilde{B}_l), quad M^{(t)} = T_r(T_c(M^{(t-1)}))

4. Hybrid Attention: CSA + HCA (§2.3)

CSA: Compressed Sparse Attention (§2.3.1)

  1. KV Compression (ratio 1/m):

    • Two streams C^a, C^b with weights Z^a, Z^b
    • Overlapped: each compressed block uses 2m entries
    • Softmax with positional biases B^a, B^b
    • C^Comp_i = sum S^a_j odot C^a_j + sum S^b_j odot C^b_j
  2. DSA on Compressed KV (lightning indexer + top-k):

    • Indexer keys compressed same way → K^IComp
    • Low-rank indexer queries from shared latent c_t^Q
    • Top-k selection: C^SprsComp_t = \{C^Comp_s mid I_{t,s} ∈ Top-k\}
    • MQA core attention + grouped output projection
  3. Sliding window KV (uncompressed) for local deps

HCA: Heavily Compressed Attention (§2.3.2)

  • Single KV stream, compression ratio 1/m' (m' gg m)
  • No overlap, no sparse selection → dense on compressed KV
  • Same MQA + grouped projection
  • Sliding window for local deps

Additional Techniques (§2.3.3)

Technique Purpose
Query/KV RMSNorm Prevent exploding logits
Partial RoPE (last 64 dims) Relative position on core outputs
Sliding window Local fine-grained deps
Attention sink Learnable sink logits per head

Efficiency (§2.3.4, Fig 1 right)

Metric V3.2 (128K) V4-Pro (1M) V4-Flash (1M)
Single-token FLOPs 100% 27% 10%
KV Cache Size 100% 10% 7%
KV Storage BF16 FP8 (non-RoPE) + BF16 (RoPE) Same
Indexer Compute FP8 FP4 FP4
Expert Weights BF16/FP8 FP4 FP4

vs BF16 GQA8 baseline (1M ctx): V4 KV cache ≈ 2%


5. Muon Optimizer (§2.4, Algorithm 1)

Algorithm

for each weight matrix W ∈ ℝ^{n×m}:
    G_t = ∇_W L_t(W_{t-1})        # Gradient
    M_t = μ M_{t-1} + G_t         # Momentum
    O'_t = HybridNewtonSchulz(μ M_t + G_t)  # Nesterov + orthogonalization
    O_t = O'_t * √max(n,m) * γ    # Rescale RMS
    W_t = W_{t-1} * (1 - ηλ) - η O_t        # Weight decay + update

Hybrid Newton-Schulz (10 steps)

Steps Coefficients (a,b,c) Purpose
1-8 (3.4445, -4.7750, 2.0315) Rapid convergence to singular values ≈ 1
9-10 (2, -1.5, 0.5) Stabilize exactly at 1
  • Pre-normalize: M_0 = M / ||M||_F
  • Avoids QK-Clip (muon_kimi) because RMSNorm on queries/KV prevents exploding logits

Optimizer Assignment

Parameter Group Optimizer
Embeddings, prediction head, RMSNorm, mHC static biases/gating AdamW
All others (attn, MoE, MTP) Muon

6. Model Variants & Specs

Spec V4-Pro V4-Flash
Total Params 1.6T 284B
Activated Params 49B 13B
Context Length 1M 1M
Experts FP4 FP4
Activated Experts 8/256 + 2 shared 6/64 + 1 shared
Layers 60 30
Hidden Dim 8192 4096
Attention Heads 64 32

7. Training

Pre-training

  • V4-Flash: 32T tokens
  • V4-Pro: 33T tokens
  • Native 1M context support from base model

Post-Training Pipeline (§3.6)

  1. Independent Expert Cultivation (per domain):

    • SFT on domain data (math, code, agent, instruction)
    • RL (GRPO) with domain-specific reward model
    • → 4+ specialized experts
  2. On-Policy Distillation (unification):

    • Unified student trained via reverse KL to teachers
    • Student acts on-policy, teachers as experts
    • Based on: online distillation (Lu et al. 2025), MiniLLM

Infrastructure Optimizations

Component Innovation
MoE Kernel Single fused kernel: comp + comm + memory overlap
TileLang DSL for kernels: productivity + efficiency
Deterministic Kernels Bitwise reproducibility train/inference
Autograd Extension Tensor-level checkpointing for fine-grained recomputation
Hybrid ZeRO + Muon Cost-effective for orthogonalization
mHC Recompute Fused kernels for Sinkhorn
2-Stage Context Parallel For compressed attention
FP4 QAT MoE experts + indexer QK path
Heterogeneous KV Cache On-disk storage + shared-prefix reuse

8. Evaluation Summary (§1, Fig 1 left)

Knowledge

  • V4-Pro-Max > open models on SimpleQA, Chinese-SimpleQA
  • Marginal lead on MMLU-Pro, HLE, GPQA
  • Gap vs Gemini-3.1-Pro narrowing

Reasoning

  • V4-Pro-Max > GPT-5.2, Gemini-3.0-Pro on standard benchmarks
  • Slightly below GPT-5.4, Gemini-3.1-Pro (~3-6 months behind)
  • V4-Flash-Max: comparable to GPT-5.2/Gemini-3.0-Pro with larger thinking budget

Agent

  • Public: ≈ Kimi-K2.6, GLM-5.1; < closed frontier
  • Internal: > Claude Sonnet 4.5; ≈ Opus 4.5

Long-Context (1M)

  • Strong on synthetic & real use cases
  • Surpasses Gemini-3.1-Pro on academic benchmarks

9. Limitations (§5)

  1. World knowledge still behind proprietary (less pre-train FLOPs)
  2. Token efficiency: longer CoT than Gemini-3.0-Pro for same quality
  3. Complex reasoning inferior to frontier

10. My Research Connections

Insight My Wiki Link
CSA/HCA → efficient ViT attention csa-hca-attention, vision-transformer-dsa-integration
mHC → stable deep residual connections mhc-hyperconnections
Muon → faster LLM training muon-optimizer
1M context → long-horizon video/agent tasks Relevant for world models
FP4 experts → quantization research quantization-awareness-training
Expert distillation → model merging Could apply to vision experts

Related Wiki Pages


Sources

  • arxiv-2606.19348: DeepSeek-V4 paper (full)