Document versionv2.0 (2026-07-27) ·Verification basis DeepSeek official HF model cards / DeepSeek API Docs / technical report (arXiv 2606.19348) Scope changev1 covered Flash only; V4 is aPro/Flash two-model family release, so both are covered together. The architecture is identical — separate write-ups would cause confusion. StatusOfficial model cards label themselves aspreview version. GA checkpoints are separate.

0. Major Corrections vs v1

# Item v1 claim Correction
1 Expert layout "256 routed + 1 shared" Expert count not disclosed on official model card. Unsourced figure removed
2 Architecture "Static token-ID→expert-ID hash table" Not among the official three upgrades (Hybrid Attention / mHC / Muon). Estimated from pre-release leaks — removed
3 Attention "Hybrid local + long-range" Precisely CSA (Compressed Sparse Attention) + HCA (Heavily Compressed Attention)
4 Flash hardware "FP8 H100 ×2 / INT4 H100 ×1" Wrong.Instruct is FP4+FP8 mixed ~158GB. With 1M context ~170–175GB →2×H200 or 4×A100. 2×H100 (160GB) is insufficient
5 Flash memory "FP8 ~500GB, INT4 141–155GB" Figures do not apply to Flash. Contradicts §2.1
6 Recommended setup "16-GPU machine" Oversized for Flash. 16 GPUs is a Pro multi-node requirement
7 Coding performance "Flash is top-tier for coding" Model card applies that label to V4-Pro-Max. Flash-Max is LiveCodeBench 91.6 (Pro-Max 93.5)
8 Chat template apply_chat_template example Model card explicitly states no Jinja template. Must use encoding_dsv4 — v1 code does not work
9 Sampling temperature 0.7 / top_p 0.9 Official recommendation is 1.0 / 1.0
10 vLLM --enforce-eager Disables CUDA graphs and sharply reduces production throughput. Removed
11 Reasoning modes Not documented Non-think / Think High / Think Max three modes are central to V4 usage. Completely missing in v1
12 DSpark Not documented Released 2026-06-27. 60–85% higher per-user generation speed vs MTP-1, lossless
13 Claude Code api.deepseek.com/v1 Wrong. Anthropic-compatible endpoint is /anthropic, model ID is deepseek-v4-pro[1m]
14 Legacy API Not documented deepseek-chat / deepseek-reasoner ended 2026-07-24 15:59 UTC (already past)
15 Korean "Benchmarks not published" No DeepSeek official Korean benchmarks, but third-party (Upstage) evaluation averages 84.9 — changes Korea adoption calculus
16 GGUF Table stated as confirmed fact Community builds, unverified vs official checkpoints. Wording softened
17 Training specs Partially documented 32T tokens / Muon optimizer / 2-stage post-training (per-domain expert training → on-policy distillation merge)
18 Output limit Not documented Max output 384K tokens. Think Max recommends context window ≥384K

1. Project Background

DeepSeek-V4 is an open-weight MoE model family released by China's DeepSeek in late April 2026. The technical report title states the design intent plainly — Towards Highly Efficient Million-Token Context Intelligence. The goal is to make 1M-token context economically viable.

Release date note: Sources disagree (April 22, 23, 24, 27). HF paper publication is April 26. This document uses "late April 2026"; confirm exact dates via DeepSeek official channels if needed.

1.1 Three Official Architecture Upgrades

Item Details
Hybrid Attention CSA + HCA combined. At 1M context, V4-Pro achieves27% single-token inference FLOPs and 10% KV cache vs V3.2
mHC Manifold-Constrained Hyper-Connections. Strengthens residual connections for more stable inter-layer signal propagation while preserving expressiveness
Muon Optimizer Improved convergence speed and training stability

CSA+HCA replaces V3.2's MLA (Multi-head Latent Attention). The idea is to bypass quadratic attention growth via compression; reducing KV cache to one-tenth is essentially the entire economic story of this model.

1.2 Training Pipeline

  • Pre-training: 32T+ high-quality tokens
  • Two-stage post-training:
    1. Individually train per-domain expert models (SFT + GRPO-based RL)
    2. Merge into a single model via on-policy distillation

This "train separately, then merge" structure likely explains V4's relatively small domain variance.

1.3 Model Lineup

Model Total Params Active Context Precision
DeepSeek-V4-Flash-Base 284B 13B 1M FP8 Mixed
DeepSeek-V4-Flash 284B 13B 1M FP4 + FP8 Mixed*
DeepSeek-V4-Pro-Base 1.6T 49B 1M FP8 Mixed
DeepSeek-V4-Pro 1.6T 49B 1M FP4 + FP8 Mixed*

* MoE expert parameters in FP4; most other weights in FP8

Derived checkpoints: DeepSeek-V4-Pro-DSpark, DeepSeek-V4-Flash-DSpark (§2.5), nvidia/DeepSeek-V4-Pro-NVFP4

1.4 Reasoning Modes (Core to V4 Usage)

Mode Characteristics Use case Response format
Non-think Fast intuitive response Daily work, low-risk decisions </think> summary
Think High Deliberate logical analysis, slower but accurate Complex problem solving, planning <think> reasoning </think> summary
Think Max Maximum reasoning extension Exploring model reasoning limits Special system prompt + <think> reasoning </think> summary
Benchmark Flash Non-Think Flash Max Pro Non-Think Pro Max
HLE 8.1 34.8 7.7 37.7
LiveCodeBench 55.2 91.6 56.8 93.5
Apex 1.0 33.0 0.4 38.3
MRCR 1M 37.5 78.7 44.7 83.5
BrowseComp – 73.2 – 83.4

The gap between Non-think and Max is 36 points on LiveCodeBench and 32 on Apex. Choosing the wrong mode costs more than switching models.

2. Strengths

2.1 Long-Context Economics — The Essence of This Model

10% KV cache and 27% inference FLOPs vs V3.2. The differentiator is not "supports 1M context" but supports 1M at a bearable cost. Flash's full 1M KV cache is ~10GB, so on top of 158GB weights it still fits on 2×H200.

2.2 MIT License — Practically the Biggest Strength

No restrictions on weights, derivative models, redistribution, or commercial use. No model-name prefix or attribution requirements. Compared to Upstage Solar License (naming/attribution obligations) or Llama-family licenses, friction for embedding in your own branded product is zero.

From an enterprise perspective this can matter more than benchmark scores. It is the only quasi-frontier class with effectively no legal review burden.

2.3 Flash Self-Hosting Fit

Flash 284B-A13B at FP4+FP8 ~158GB; with 1M context KV and runtime overhead ~170–175GB total. 2×H200 (282GB) fits comfortably. Practitioners commonly estimate ~85–95% of Pro quality at far lower infrastructure.

2.4 Overwhelming API Pricing

Model Input (cache miss) Output Input (cache hit)
DeepSeek-V4-Pro $0.435 / 1M $0.87 / 1M $0.003625 / 1M
DeepSeek-V4-Flash $0.14 / 1M $0.28 / 1M $0.0028 / 1M

As of May 22, 2026, V4-Pro's 75% discount became the standing price. Cache-hit input at $0.003625/1M is especially important — input cost effectively vanishes for agent workloads reusing the same system prompt.

2.5 DSpark — Serving Optimization (2026-06-27)

Not a new model — draft modules attached to existing V4 weights. Official cards state "not a new model."

Item Details
Method Parallel draft backbone + small sequential head, confidence head and load-aware scheduler
Effect 60–85% higher per-user generation speed vs MTP-1, lossless output
Throughput +51% / +52% total throughput at 80 tok/s/user (Flash) · 35 tok/s/user (Pro)
High-speed case +661% / +406% at 120 tok/s/user · 50 tok/s/user
Checkpoints DeepSeek-V4-Pro-DSpark, DeepSeek-V4-Flash-DSpark (MIT)
Code DeepSpec (github.com/deepseek-ai/DeepSpec, MIT)

Note: 60–85% is vs DeepSeek's own MTP-1, not naive decoding. No guarantee of reproduction on other serving stacks; no independent replication reports as of early July 2026.

2.6 Agent Tool Ecosystem Integration

Official Anthropic-compatible endpoint connects Claude Code, OpenCode, Copilot Chat, Cline, etc. without a proxy. The awesome-deepseek-agent repo documents official setup for 20 coding tools.

2.7 Base Model Release

V4-Pro-Base and V4-Flash-Base enable continued pretraining. A decisive difference from instruct-only releases; if you plan serious domain-specific models, this alone can be the selection reason.

3. Weaknesses and Limitations

3.1 Pro Is Effectively Cluster-Only

Configuration Requirement
V4-Pro weights ~862GB (HF listing) / vLLM recommended recipe ~960GB footprint
Single node 8×H200 141GB (1,128GB) or B300 8-GPU node
Multi-node 16×H100 80GB across 2 nodes + NVLink/InfiniBand
8×H100 80GB Not viable (640GB insufficient)

Quantization does not fully solve this. Even at Q4 ~430GB, adding 1M KV exceeds 8×H100 again. No quantization level makes Pro a workstation model — only a slightly smaller cluster model.

3.2 No Chat Template — Biggest Integration Trap

Model cards state: this release does not include a Jinja chat template. Use Python scripts in the encoding folder (encoding_dsv4) to encode messages and parse output.

If your pipeline depends on tokenizer.apply_chat_template(), that code will not work. vLLM/SGLang handle this internally, but Transformers direct use or custom serving requires implementing the encoding layer yourself.

3.3 Asymmetric Official Multilingual Information

DeepSeek official benchmarks center on English and Chinese (C-Eval, CMMLU, Chinese-SimpleQA). No Korea/Japan-specific benchmarks published. Indirect evidence: Base MultiLoKo (multilingual) improved V3.2 38.7 → Flash 42.2 → Pro 51.1.

This does not mean "cannot do Korean." See §5.

3.4 Serving Stack Constraints

Stack Status
vLLM Native support v0.22.0, production hardening v0.23.0
SGLang Day-0 official support. MegaMoE is Blackwell-only (B200/B300/GB200/GB300)
TGI Not supported at preview time
Ollama / llama.cpp Community GGUF only, unverified vs official checkpoints

Also, as of May 2026, multiple reports of vLLM Inductor compile-path crashes on RTX Pro 6000 Blackwell. Verify current vLLM/driver state before buying that hardware — memory may be sufficient while software path is unstable.

3.5 Flash Performance Ceiling

Flash-Max approaches Pro-level reasoning with large thinking budget, but lags Pro on pure knowledge work and the most complex agent workflows — acknowledged on the model card.

SimpleQA-Verified: Flash-Max 34.1 vs Pro-Max 57.9 — a 23.8-point gap. Flash is risky when factual accuracy matters.

3.6 Geopolitical / Regulatory Risk (Hosted API Only)

DeepSeek hosted services/apps/APIare banned or restricted on government devices in many countries. Korea's PIPC suspended the app in 2025 over cross-border data transfer and issued corrective orders; many ministries, public agencies, and financial firms blocked access. The US federal government and many states banned government-device use; legislation introduced. Japan, Australia, Taiwan, Italy, and others took similar measures.Key distinction: These rules target hosted services with data sent to China, not running MIT-licensed weights on your own infrastructure. This distinction underpins §9–11 business scenarios.

Some organizations may exclude "China-origin models" in procurement regardless — verify in advance.

3.7 Preview Status

Model cards label themselves preview. Quality may change at GA — use preview baselines as evaluation reference lines, not final deliverable quality guarantees.

4. Benchmarks (Official Model Card)

4.1 V4-Pro-Max vs Frontier Models

Benchmark Opus-4.6 Max GPT-5.4 xHigh Gemini-3.1-Pro High K2.6 Thinking GLM-5.1 Thinking DS-V4-Pro Max
Knowledge & Reasoning
MMLU-Pro 89.1 87.5 91.0 87.1 86.0 87.5
SimpleQA-Verified 46.2 45.3 75.6 36.9 38.1 57.9
Chinese-SimpleQA 76.4 76.8 85.9 75.9 75.0 84.4
GPQA Diamond 91.3 93.0 94.3 90.5 86.2 90.1
HLE 40.0 39.8 44.4 36.4 34.7 37.7
LiveCodeBench 88.8 – 91.7 89.6 – 93.5
Codeforces (Rating) – 3168 3052 – – 3206
HMMT 2026 Feb 96.2 97.7 94.7 92.7 89.4 95.2
IMOAnswerBench 75.3 91.4 81.0 86.0 83.8 89.8
Apex 34.5 54.1 60.9 24.0 11.5 38.3
Apex Shortlist 85.9 78.1 89.1 75.5 72.4 90.2
Long Context
MRCR 1M 92.9 – 76.3 – – 83.5
CorpusQA 1M 71.7 – 53.8 – – 62.0
Agent
Terminal Bench 2.0 65.4 75.1 68.5 66.7 63.5 67.9
SWE Verified 80.8 – 80.6 80.2 – 80.6
SWE Pro 57.3 57.7 54.2 58.6 58.4 55.4
SWE Multilingual 77.5 – – 76.7 73.3 76.2
BrowseComp 83.7 82.7 85.9 83.2 79.3 83.4
HLE w/ tools 53.1 52.0 51.6 54.0 50.4 48.2
GDPval-AA (Elo) 1619 1674 1314 1482 1535 1554
MCPAtlas Public 73.8 67.2 69.2 66.6 71.8 73.6
Toolathlon 47.2 54.6 48.8 50.0 40.7 51.8
  • MCPAtlas 73.6 is effectively tied with Opus-4.6 (73.8). Rare case of open weight matching closed on MCP agents.
  • SimpleQA-Verified 57.9 trails Gemini (75.6) badly. Unsuitable for factual lookup without RAG.
  • Long-context measured scores underwhelm.MRCR 1M 83.5, CorpusQA 1M 62.0 vs Opus-4.6 (92.9 / 71.7).Can fit 1M" ≠ "reads accurately.
  • Apex 38.3 trails Gemini (60.9) and GPT-5.4 (54.1) substantially.

4.2 Mode Comparison (Flash vs Pro)

Benchmark Flash Non-Think Flash High Flash Max Pro Non-Think Pro High Pro Max
MMLU-Pro 83.0 86.4 86.2 82.9 87.1 87.5
SimpleQA-Verified 23.1 28.9 34.1 45.0 46.2 57.9
Chinese-SimpleQA 71.5 73.2 78.9 75.8 77.7 84.4
GPQA Diamond 71.2 87.4 88.1 72.9 89.1 90.1
HLE 8.1 29.4 34.8 7.7 34.5 37.7
LiveCodeBench 55.2 88.4 91.6 56.8 89.8 93.5
Codeforces – 2816 3052 – 2919 3206
HMMT 2026 Feb 40.8 91.9 94.8 31.7 94.0 95.2
Apex 1.0 19.1 33.0 0.4 27.4 38.3
MRCR 1M 37.5 76.9 78.7 44.7 83.3 83.5
CorpusQA 1M 15.5 59.3 60.5 35.6 56.5 62.0
Terminal Bench 2.0 49.1 56.6 56.9 59.1 63.3 67.9
SWE Verified 73.7 78.6 79.0 73.6 79.4 80.6
SWE Pro 49.1 52.3 52.6 52.1 54.4 55.4
BrowseComp – 53.5 73.2 – 80.4 83.4
MCPAtlas 64.0 67.4 69.0 69.4 74.2 73.6
Toolathlon 40.7 43.5 47.8 46.3 49.0 51.8
  • MCPAtlas: Pro High (74.2) > Pro Max (73.6).Excessive reasoning hurts tool calling.Fix agent orchestration to High.-Do not use Non-think for coding, math, or agents. LiveCodeBench 55.2, Apex 1.0, HLE 8.1 — effectively useless. Non-think is for classification, summarization, routing only.
  • SWE Verified: Flash High 78.6 vs Pro Max 80.6 — 2-point gap. Flash is enough for code-fix tasks.

4.3 Base Models (Pre-training Quality)

Benchmark V3.2-Base V4-Flash-Base V4-Pro-Base
MMLU (5-shot) 87.8 88.7 90.1
MMLU-Pro (5-shot) 65.5 68.3 73.5
MMMLU (5-shot) 87.9 88.8 90.3
C-Eval (5-shot) 90.4 92.1 93.1
CMMLU (5-shot) 88.9 90.4 90.8
MultiLoKo (multilingual) 38.7 42.2 51.1
Simple-QA verified 28.3 30.1 55.2
FACTS Parametric 27.1 33.9 62.6
SuperGPQA 45.0 46.5 53.9
HumanEval 62.8 69.5 76.8
BigCodeBench 63.9 56.8 59.2
MATH 60.5 57.4 64.5
LongBench-V2 40.2 44.7 51.5

Caveat: V4-Flash-Base trails V3.2-Base on BigCodeBench and MATH — cost of active params 37B → 13B. Uncritical Flash-as-V3.2-replacement can regress.

5. Korean and Japanese Performance — No Official Data, Third-Party Data Exists

DeepSeek has not published Korean benchmarks. However, Upstage evaluated DeepSeek-V4-Flash as a comparator when announcing Solar Open 2 (2026-07-22). As a competitor's table meant to highlight their model, there is little incentive to inflate V4-Flash.

Korean Benchmark DeepSeek-V4-Flash Solar Open 2 GPT-5.4 mini Claude Haiku 4.5
KMMLU-Pro 78.9 78.4 78.1 67.9
CLIcK 89.2 90.7 89.6 53.5
HAE-RAE v1.1 73.1 73.8 69.4 38.5
Ko-AIME'25 98.0 97.7 90.7 81.7
HRM8K 93.4 92.2 91.3 90.6
KBank-MMLU (finance) 79.5 80.8 79.0 68.9
KBL (legal) 72.8 75.5 75.3 69.9
KorMedMCQA 94.1 93.0 94.2 87.0
Ko-GDPval 85.0 86.8 59.4 68.3
Korean average 84.9 85.4 80.8 69.6

Also, DeepSeek-V4-Pro (1.6T) scored 86.91 on Ko-GDPval, narrowly ahead of Solar Open 2 (86.75).

The conclusion flips.v1's "insufficient Korean performance info → fine-tuning required for Korea" is wrong.DeepSeek-V4-Flash is within 0.5 points of Korea-optimized models on Korean average.Korean math (Ko-AIME 98.0, HRM8K 93.4) and general knowledge (KMMLU-Pro 78.9) actually lead.Japanesestill has no public data. Nejumi leaderboard or self-run JMMLU needed.Korean practical notes- Tokenizer is not Korean-optimized —higher token cost than Korean-specialized models. Measure cost even if quality is comparable

  • Korean document formats and administrative idioms may be underrepresented. Ko-GDPval 85.0 is still strong; few-shot in-house format injection is essential
  • Response bias reported on Chinese political/historical topics — validate if your domain touches them

6. Competing Projects

6.1 Solar Open 2 (Upstage, 250B-A15B)

Item DeepSeek-V4-Flash Solar Open 2
Developer DeepSeek (China) Upstage (Korea)
Total / Active 284B / 13B 250B / 15B
Release Late April 2026 2026-07-22
Context 1M 1M
License MIT (unrestricted) Upstage Solar License (naming/attribution)
Base released Yes No
Hardware 2×H200 (FP4+FP8) 4×H200 (BF16) / 2×H200 (quantized)
Korean average 84.9 85.4
MMLU-Pro 85.9 86.2
LiveCodeBench 92.3 92.4
SWE-Bench Verified 73.8 70.4
MCP-Atlas 58.2 58.2
Terminal Bench Hard 34.1 28.3
τ³ (banking) 22.3 19.6
APEX-Agents 13.2 16.6

Korean enterprise choice splits on other axes — domestic procurement → Solar; license freedom and fine-tuning flexibility → DeepSeek; China-exclusion policy → Solar.

6.2 Kimi K3 (Moonshot AI, 2.8T)

Item DeepSeek-V4-Pro Kimi K3
Total params 1.6T 2.8T
Active 49B 16 experts active (of 896)
Release Late April 2026 API 7/16, weights 2026-07-27
License MIT Modified MIT
Multimodal Text Native vision
Context 1M 1M
API price $0.435 / $0.87 $3 / $15
Reasoning 3 selectable modes thinking always on

Per model card vs K2.6 Thinking: V4-Pro-Max leads LiveCodeBench (93.5 vs 89.6), MCPAtlas (73.6 vs 66.6); trails SWE Pro (55.4 vs 58.6), HLE w/tools (48.2 vs 54.0). K3 is one generation ahead — re-evaluate after weight release.

Price gap is decisive. V4-Pro is ~1/7–1/17 of K3. For most production workloads not needing absolute frontier peak, V4-Pro economics dominate.

6.3 V4-Pro vs V4-Flash Selection

Condition Recommendation
Factual accuracy matters (SimpleQA 57.9 vs 34.1) Pro
Code fix / test pass (SWE 80.6 vs 79.0) Flash — 1.6-point gap, 5× infrastructure
Long-context accuracy (MRCR 83.5 vs 78.7) Pro
MCP tool calling (74.2 vs 69.0) Pro High
Browsing agent (83.4 vs 73.2) Pro — 10-point gap
Self-hosting budget constrained Flash
Classification / routing / summarization Flash Non-think

General rule: Default Flash High; route SimpleQA·BrowseComp-type tasks to Pro. DeepSeek's official Claude Code recipe uses exactly this (main Pro, sub-agent Flash).

7. Getting Started

7.1 Hardware Requirements

Model Precision Weights Total VRAM budget Configuration
V4-Flash FP4+FP8 ~158GB ~170–175GB 2×H200 or 4×A100 80GB
V4-Flash Community INT4 ~90GB – 4×RTX 4090 (unverified)
V4-Pro FP4+FP8 ~862GB ~960GB 8×H200 141GB single node or B300 8-GPU
V4-Pro Multi-node – – 16×H100 80GB 2-node + IB
  • System RAM: 256GB+ for Flash; NVMe storage 500GB+
  • vLLM tensor parallel optimal at powers of 2 (1/2/4/8). 2×A100 (160GB) insufficient for 1M context — prefer 4×A100
  • V4-Pro does not fit on 8×H100 80GB (640GB)### 7.2 Chat Encoding (Required Prerequisite)This release has no Jinja chat template. Use scripts in the encoding folder.
from encoding_dsv4 import encode_messages, parse_message_from_completion_text
import transformers

messages = [
    {"role": "user", "content": "hello"},
    {"role": "assistant", "content": "Hello! I am DeepSeek.", "reasoning_content": "thinking..."},
    {"role": "user", "content": "1+1=?"},
]

# messages -> string
prompt = encode_messages(messages, thinking_mode="thinking")

# string -> tokens
tokenizer = transformers.AutoTokenizer.from_pretrained("deepseek-ai/DeepSeek-V4-Pro")
tokens = tokenizer.encode(prompt)

In multi-turn dialogue, preserve prior turns' reasoning_content.

7.3 vLLM Serving (Recommended)

# vLLM v0.23.0+ recommended (native from v0.22.0)
pip install -U vllm

# V4-Flash: 2×H200, 1M context
vllm serve deepseek-ai/DeepSeek-V4-Flash \
  --served-model-name deepseek-v4-flash \
  --tensor-parallel-size 2 \
  --max-model-len 1048576 \
  --enable-expert-parallel

# Conservative start (4×A100, 128K context)
vllm serve deepseek-ai/DeepSeek-V4-Flash \
  --tensor-parallel-size 4 \
  --max-model-len 131072

# V4-Pro: 8×H200
vllm serve deepseek-ai/DeepSeek-V4-Pro \
  --served-model-name deepseek-v4-pro \
  --tensor-parallel-size 8 \
  --enable-expert-parallel \
  --max-model-len 393216   # Think Max recommends ≥384K

v1's --enforce-eager was removed. It disables CUDA graphs and sharply cuts production throughput.

7.4 SGLang Serving

pip install sglang

python3 -m sglang.launch_server \
    --model-path deepseek-ai/DeepSeek-V4-Flash \
    --host 0.0.0.0 --port 30000 \
    --tp 2 \
    --context-length 1048576

MegaMoE backend (higher throughput with negligible accuracy loss, Pro GPQA ~89.5) is Blackwell-only (B200/B300/GB200/GB300). Not available on Hopper (H100/H200).

7.5 DSpark Deployment

# Existing weights + draft module. Not a separate model
vllm serve deepseek-ai/DeepSeek-V4-Flash-DSpark \
  --tensor-parallel-size 2 \
  --max-model-len 1048576 \
  --enable-expert-parallel

Same output distribution — throughput gain without quality regression. Still measure whether 60–85% reproduces on your stack.

7.6 Sampling Parameters

Parameter Official recommendation Notes
temperature 1.0 v1's 0.7 was wrong
top_p 1.0 v1's 0.9 was wrong
max output up to 384K
context (Think Max) ≥384K Truncation mid-reasoning if insufficient

7.7 API Usage (Without Self-Hosting)

curl https://api.deepseek.com/v1/chat/completions \
  -H "Content-Type: application/json" \
  -H "Authorization: Bearer $DEEPSEEK_API_KEY" \
  -d '{
    "model": "deepseek-v4-flash",
    "messages": [{"role": "user", "content": "Explain this"}],
    "thinking": true
  }'
  • Supports both OpenAI ChatCompletions and Anthropic formats
  • Thinking mode controlled via thinking parameter, not model ID
  • Legacy deepseek-chat / deepseek-reasoner ended 2026-07-24 15:59 UTC — returns HTTP 400. Switch to deepseek-v4-flash / deepseek-v4-pro
  • Hosted API sends data to China. Do not use for sensitive information (§3.6)

7.8 Claude Code Integration (Official)

export ANTHROPIC_BASE_URL=https://api.deepseek.com/anthropic
export ANTHROPIC_AUTH_TOKEN=<your DeepSeek API Key>
export ANTHROPIC_MODEL=deepseek-v4-pro[1m]
export ANTHROPIC_DEFAULT_OPUS_MODEL=deepseek-v4-pro[1m]
export ANTHROPIC_DEFAULT_SONNET_MODEL=deepseek-v4-pro[1m]
export ANTHROPIC_DEFAULT_HAIKU_MODEL=deepseek-v4-flash
export CLAUDE_CODE_SUBAGENT_MODEL=deepseek-v4-flash
export CLAUDE_CODE_EFFORT_LEVEL=max
cd /path/to/project && claude

v1's api.deepseek.com/v1 is wrong. Anthropic path is /anthropic; [1m] suffix is DeepSeek's 1M context modifier — keep as documented. Gateways may need suffix normalization for billing.

For self-hosted vLLM, replace the URL with your internal endpoint — standard path in regulated environments.

7.9 GGUF / Local Run (Unverified)

Community GGUF exists for llama.cpp/Ollama but is unverified vs official checkpoints. Experiments only; production should use vLLM or SGLang. Mainline support varies — verify current state before adoption.

8. Key Use Cases

8.1 Full-Repository Code Agent — Strongest Area

LiveCodeBench 93.5, Codeforces 3206 beat closed frontiers. With 1M context, loading an entire monolithic legacy codebase for refactoring is viable.

# Claude Code on internal vLLM
export ANTHROPIC_BASE_URL=http://gpu-node.internal:8000
export ANTHROPIC_AUTH_TOKEN=dummy
export ANTHROPIC_MODEL=deepseek-v4-pro
export ANTHROPIC_DEFAULT_HAIKU_MODEL=deepseek-v4-flash
export CLAUDE_CODE_SUBAGENT_MODEL=deepseek-v4-flash
cd ~/legacy-erp && claude

Cost design is key. Orchestration on Pro, sub-agents on Flash — minimal quality loss, large cost drop. SWE Verified Flash High 78.6 vs Pro Max 80.6 — no reason for Pro on sub-agents.

8.2 1M Context Document Analysis — With Verification

from openai import OpenAI
client = OpenAI(api_key="EMPTY", base_url="http://localhost:8000/v1")

resp = client.chat.completions.create(
    model="deepseek-v4-pro",
    messages=[{"role": "user", "content": f"""
Below are the full main contract, three supplementary agreements, and internal standard contract guidelines.

{full_documents}

1. List all conflicting clauses across documents with clause numbers
2. Clauses violating standard guidelines and reasons
3. Risk grade per clause and negotiation priority

Do not speculate without evidence; cite original clause text.
"""}],
    temperature=1.0, top_p=1.0, max_tokens=131072,
)

Must-state limits: MRCR 1M 83.5, CorpusQA 1M 62.0 — below Opus-4.6 (92.9 / 71.7). Fitting in context does not guarantee accurate reading.

Practical guidance:

  • Operate at 200–400K, not full 1M
  • Cross-verify critical clauses with two independent queries
  • Run your own needle-in-haystack tests first

8.3 MCP-Based Agent Orchestration

MCPAtlas Public 73.6 ties Opus-4.6 (73.8) — top among open weights.

resp = client.chat.completions.create(
    model="deepseek-v4-pro",
    messages=[...],
    tools=mcp_tools,
    extra_body={"thinking": True},   # Think High
)

Use Think High. MCPAtlas Pro High 74.2 > Pro Max 73.6 — max reasoning hurts. Excessive deliberation increases tool-call errors.

8.4 Factual Lookup Workloads — Prohibited or RAG Required

SimpleQA-Verified: Pro-Max 57.9, Flash-Max 34.1 — far below Gemini-3.1-Pro (75.6).

Flash factual Q&A may be wrong two-thirds of the time. Internal knowledge Q&A, regulation lookup, customer support must use RAG and system prompts that refuse unsupported answers.

SYSTEM = """Never answer content not grounded in provided documents.
If no evidence exists, respond 'Information not found.'
Cite source document and page for every stated fact."""

8.5 Domain-Specific Continued Pretraining

Base model release enables this scenario. MIT license — no derivative naming or attribution obligations.

V4-Flash-Base → continued pretraining on domestic finance rules/terms → own-brand model
V4-Flash-Base → continued pretraining on Japanese corpus → Japanese-specialized model
V4-Flash-Base → internal codebase + technical docs → org-specific coding model

284B-A13B is near the practical upper bound for continued pretraining. 1.6T Pro-Base training infrastructure is effectively national-scale.

8.6 Cost-Optimized Routing

def route(task_type: str) -> dict:
    return {
        "classify":   {"model": "deepseek-v4-flash", "thinking": False},
        "summarize":  {"model": "deepseek-v4-flash", "thinking": False},
        "code_fix":   {"model": "deepseek-v4-flash", "thinking": True},
        "code_arch":  {"model": "deepseek-v4-pro",   "thinking": True},
        "tool_agent": {"model": "deepseek-v4-pro",   "thinking": True},   # not Max
        "fact_query": {"model": "deepseek-v4-pro",   "thinking": True},   # + RAG
        "browse":     {"model": "deepseek-v4-pro",   "thinking": True},
    }[task_type]

Cache-hit input at $0.003625/1M — fix system prompts and maximize prefix caching to nearly eliminate input cost. Largest effect on agent workloads.

9. Korea Market Adoption Scenarios

9.0 Prerequisite: Understand Regulatory Scope Precisely

In 2025 Korea's PIPC suspended DeepSeek's app over cross-border transfer and issued corrective orders; many ministries (Foreign Affairs, Defense, MOTIE), public agencies (KHNP, KEPCO KPS), and major financial firms blocked access. Kakao, LINE Yahoo, and other private IT firms banned business use.

These measures target hosted services.Running MIT weights on in-house GPUs sends no data abroad — the regulatory logic does not apply.Therefore the only legitimate Korea adoption path is self-hosting. Treat API use as off the table.

Also acknowledge reality: some orgs penalize "China-origin model" in procurement regardless of technical validation — especially public, defense, and parts of finance. Solar Open 2 is the alternative; performance gap is ~0.5 points.

9.1 Software Development Organizations — Top Priority

Item Details
Target Game studios, SI, platforms, startup dev teams
Rationale LiveCodeBench 93.5 (#1), Codeforces 3206, SWE Verified 80.6, MIT license
Problem No external source export, commercial assistant license cost, missing legacy documentation
Setup In-house 2×H200 Flash + Claude Code/OpenCode. Pro on cloud GPU as needed
KPI PR lead time, review findings, test coverage
Strength No regulatory gray zone. No PII in code, self-hosted, MIT license

Lowest political friction and highest confidence ROI for DeepSeek-V4 in Korea.

9.2 Financial Sector — Self-Hosted Back Office

Item Details
Rationale KBank-MMLU 79.5, KBL 72.8, τ³ banking 22.3 (vs Solar 19.6), air-gapped fit
Setup Internal 2×H200 Flash, fully blocked external comms
Output Regulation Q&A (RAG required), review report drafts, code review
Risk SimpleQA 34.1 (Flash). No finance regulation Q&A without RAG
Reality Many banks already block DeepSeek access. Self-hosting still faces high internal approval bar. Pre-align with InfoSec: "running weights locally ≠ cross-border data transfer"

9.3 Manufacturing / Logistics — On-Premises Docs and Code

Item Details
Rationale 1M context, MIT license for own-brand embedding
Use PLC/MES legacy code analysis, technical doc multilingualization (KO/EN/ZH), spec cross-check
Note Chinese capability (C-Eval 93.1, CMMLU 90.8) is a real asset for firms with China operations. Solar Open 2 lacks Chinese
Risk High IP sensitivity — fully closed network required

9.4 Public / Defense — Not Recommended

Technically solvable via self-hosting, but procurement policy and NIS security review make China-origin models nearly impossible. Government requires NIS security review for AI in public IT projects.

Use Solar Open 2 or other domestic models.

9.5 Korea Scenario Priority

Rank Segment Performance Regulatory friction Overall
1 Dev org coding agent Very high Low Strongly recommended
2 Manufacturing on-prem (China-linked) High Medium Recommended
3 Domain derivative model dev High (Base released) Low Recommended
4 Finance back office Medium High Conditional
5 Public / defense – Very high Not recommended

10. Japan Market Adoption Scenarios

10.0 Prerequisite: Unverified Japanese + On-Premises Preference

Japan structurally favors on-premises — MIT open weights fit well in principle. No public Japanese performance evidence. MultiLoKo 51.1 (Pro Base) is only indirect signal.

Japan is among countries reviewing DeepSeek regulation; LINE Yahoo banned employee business use.

Required pre-work

  1. Self-measure on W&B Nejumi leaderboard basis
  2. Self-measure JMMLU / JCommonsenseQA
  3. Measure Japanese token efficiency (Chinese-optimized tokenizer may help kanji, hurt kana)
  4. PoC on keigo and business idiom accuracy

Do not propose to Japanese customers without this validation.

10.1 Japan IT / SIer Coding Agent — Top Priority

Item Details
Rationale Code is language-neutral — only use case bypassing Japanese validation
Target SIers, internal IS departments, SaaS vendors
Problem COBOL/legacy modernization, developer shortage, no external code export
Setup On-prem 2×H200 Flash + Claude Code
Strength SWE Multilingual 76.2 — strong on multilingual codebases including Japanese comments

Only scenario immediately actionable in Japan.

10.2 Japanese-Specialized Derivative Model — Mid/Long-Term Strategy

Base release + MIT license creates the opportunity.Same strategy as Upstage's Karakuri and Syn Pro (31B) on Solar for Nejumi #1 — executable on DeepSeek-V4-Flash-Basewithout license constraints. Solar License requires "Solar" prefix and "Built with Solar"; MIT requires nothing.

Stage Details
1 Continued pretraining V4-Flash-Base on Japanese corpus
2 SFT on Japanese business docs and keigo
3 Deploy under own brand (naming free)
Condition 284B CPT needs substantial compute — consider Japan partners or SoftBank AI Cloud

10.3 Japan Manufacturing — Conditional

Item Details
Rationale On-prem fit, 1M context, Chinese for China factory coordination
Risk Japanese manufacturing demands extreme document-format compliance — risky without Japanese validation
Approach Start with code/design data analysis; Japanese doc generation after validation

10.4 Japan Finance / Public — Not Recommended

Strong caution on China-origin models; long contract cycles. Low realism except joint derivative with Japan partner (§10.2).

10.5 Japan Scenario Priority

Rank Segment Japanese dependency Overall
1 Coding agent Low Start immediately
2 Japanese derivative model – (direct improvement) Mid/long-term top opportunity
3 Manufacturing design/code analysis Low Recommended
4 Japanese document work High After validation
5 Finance / public High Not recommended

11. US Market Adoption Scenarios

11.0 Prerequisite: Most Complex Regulatory Landscape

Three layers in the US:

Layer Status
General consumers No nationwide ban
Federal / state / defense / intelligence Banned on government devices/systems. Commerce, Navy, NASA, Texas, New York, Virginia, Tennessee, etc.
Government contractors May be bound by contract terms
Private enterprises Own policies

Federal legislation such as No DeepSeek on Government Devices Act (H.R.1121) progressed.

Decisive distinction: Rules target "DeepSeek application or successor applications/services" because data is sent to China servers. Running MIT weights on US infrastructure involves no data transfer.NVIDIA ships nvidia/DeepSeek-V4-Pro-NVFP4 and published on build.nvidia.com; Runpod, Spheron, and other US GPU clouds provide deployment guides.Running V4 on US infrastructure is already standard practice.

11.1 AI Startups — Changes Cost Structure

Item Details
Target Seed–Series A, frontier API cost pressuring burn rate
Rationale V4-Pro ~34× cheaper input, ~86× output vs Opus. MIT license
Setup API initially → self-host after PMF
Caution Enterprise diligence will ask "China model?" Prepare self-hosting architecture diagram
Strength MIT — no legal friction on fine-tuning, redistribution, own branding

11.2 Developer Tools / Coding SaaS

Item Details
Rationale LiveCodeBench 93.5, Codeforces 3206, SWE Verified 80.6
Economics Coding agents burn extreme tokens — 34–86× price gap creates margin
Setup 8×H200 (Pro) or 2×H200 (Flash) in own VPC
Example Morph already serves V4-Flash bf16 unquantized
Risk Many serverless hosts quantize activations to fp8 — output differs from reference weights. Self-serve if quality consistency matters

11.3 Regulated Industries (Healthcare / Finance) — VPC Deployment

Item Details
Rationale HIPAA / SOC2 / GLBA require data cannot leave
Setup Self-deploy in AWS/Azure/GCP VPC, block external egress
Economics p5.48xlarge on-demand ~$55/h, 1-year reserved ~$33/h
Cold calculation API break-even at~3–4 billion tokens/day. Single 8×H100 node cannot physically handle that
Conclusion Self-hosting is for data sovereignty / compliance / fine-tuning, not cost. Cost-only justification fails diligence

11.4 Federal Contractors / Public — Contract Review Required

Self-hosting is technically possible, but contracts may exclude China-origin software. Whether weights count as software varies by contract.

Do not proceed without legal review. If proceeding:

  • Written pre-inquiry to Contracting Officer (CO)
  • Document model origin, deployment architecture, data flows
  • Parallel evaluation of alternatives (Llama, Mistral, GPT-OSS, etc.)

11.5 US Scenario Priority

Rank Segment Economics Regulatory friction Overall
1 Coding SaaS / dev tools Very high Low Strongly recommended
2 AI startup backend Very high Medium (diligence) Recommended
3 Regulated industry VPC Low (sovereignty goal) Medium Conditional
4 Federal contractor – High Legal review required
5 Government direct – Prohibited Not possible

11.6 Three-Country Comparison Summary

Axis Korea Japan US
Regulatory nature Cross-border PII + public blocks Review stage + private bans Government/contractor explicit bans
Self-hosting legitimacy Holds Holds Holds (established practice)
Language evidence Third-party sufficient (84.9) None (validation required) Not an issue
Best entry Dev org coding Coding + derivative dev Coding SaaS
Biggest obstacle China-exclusion sentiment Japanese unverified Enterprise diligence
Alternative Solar Open 2 (comparable) Syn Pro etc. Llama, Mistral, GPT-OSS

Coding agent is #1 in all three markets — no language validation needed, strongest performance evidence (LiveCodeBench #1), minimal PII/regulatory friction.

12. Adoption Decision Checklist

12.1 Technical Validation

  • Measure Flash vs Pro quality gap on real tasks — try to disprove "Flash is enough for most"
  • Measure quality / cost / latency across modes (Non-think / High / Max) — biggest lever
  • Own needle-in-haystack at 1M context (meaning of MRCR 83.5 / CorpusQA 62.0)
  • SimpleQA mitigation: RAG pipeline and refusal of unsupported answers
  • encoding_dsv4 integration — audit all apply_chat_template dependencies
  • Measure Korean token efficiency (cost may differ despite comparable quality)
  • For Japanese: self-evaluate Nejumi / JMMLU (required)
  • Measure throughput before/after DSpark
  • Check Chinese political/historical bias if applicable

12.2 Infrastructure

  • vLLM v0.23.0+. SGLang MegaMoE Blackwell-only
  • RTX Pro 6000 Blackwell: current Inductor compile issue status
  • Pro: confirm 8×H200 single-node availability (8×H100 not viable)
  • Serverless hosts: check fp8 activation quantization — output differs from reference
  • Preview → GA re-evaluation plan
  • Cache-hit optimization design (prefix caching)

12.3 Governance / Legal

  • Decide hosted API prohibition — sensitive data has no alternative to self-hosting
  • Check org procurement China-software exclusion
  • (US) Federal contract clause review, CO pre-inquiry
  • (Korea) Pre-align InfoSec: "local weights ≠ cross-border transfer"
  • Enterprise customer diligence materials (deployment architecture, data flow)
  • Include MIT license copy (only obligation)

12.4 Cost

  • Recalculate self-hosting break-even — ~3–4B tokens/day vs API. Most orgs never reach it
  • Document real self-hosting reason: sovereignty / compliance / fine-tuning / latency
  • Measure per-mode token use and set routing policy

13. License

MIT License. Applies to repository and model weights.

Item Details
Commercial use Free
Modification / fine-tuning Free
Redistribution Free
Derivative model naming No constraints
Attribution None required
Only obligation Include license and copyright notice copy

Most permissive vs Upstage Solar License (prefix/attribution), Llama Community License (user limits/attribution), Modified MIT (Kimi K3). Alone sufficient reason if embedding in own-branded product.

14. Citation

@misc{deepseekai2026deepseekv4,
      title={DeepSeek-V4: Towards Highly Efficient Million-Token Context Intelligence},
      author={DeepSeek-AI},
      year={2026},
}

15. References

Category Link
V4-Pro model card https://huggingface.co/deepseek-ai/DeepSeek-V4-Pro
V4-Flash model card https://huggingface.co/deepseek-ai/DeepSeek-V4-Flash
V4-Pro-Base https://huggingface.co/deepseek-ai/DeepSeek-V4-Pro-Base
V4-Flash-Base https://huggingface.co/deepseek-ai/DeepSeek-V4-Flash-Base
DSpark checkpoint https://huggingface.co/deepseek-ai/DeepSeek-V4-Pro-DSpark
DSpark (Flash) https://huggingface.co/deepseek-ai/DeepSeek-V4-Flash-DSpark
NVIDIA NVFP4 quantization https://huggingface.co/nvidia/DeepSeek-V4-Pro-NVFP4
Technical report https://arxiv.org/abs/2606.19348
API documentation https://api-docs.deepseek.com/
Claude Code integration (official) https://api-docs.deepseek.com/quick_start/agent_integrations/claude_code/
Coding agent integration (official) https://api-docs.deepseek.com/guides/coding_agents/
SGLang deployment guide https://lmsysorg.mintlify.app/cookbook/autoregressive/DeepSeek/DeepSeek-V4
DeepSpec (spec decoding) https://github.com/deepseek-ai/DeepSpec

Benchmark figures are from DeepSeek official model cards. Korean data is third-party evaluation published by Upstage at Solar Open 2 announcement — not DeepSeek official. Model cards state preview status; base adoption decisions on self-reproduction evaluation.