DeepSeek-V4百万上下文开源大模型性能如何？值得用吗？

引言

DeepSeek-V4是DeepSeek推出的新一代大语言模型系列预览版，拥有百万字超长上下文窗口，在Agent能力、世界知识与推理性能方面达到开源领域领先水平。模型包含 deepseek-v4-pro 和 deepseek-v4-flash 两个版本，分别定位高性能与经济高效，均已开源并提供API服务，支持非思考与思考双模式，为长文本处理与智能体应用提供普惠化基础设施。

DeepSeek-V4 is a preview of the next-generation large language model series from DeepSeek, featuring a million-character ultra-long context window and achieving leading open-source performance in agent capabilities, world knowledge, and reasoning. The series includes two versions: deepseek-v4-pro (high performance) and deepseek-v4-flash (cost-efficient), both open-sourced and available via API. It supports both non-thinking and thinking modes, providing affordable infrastructure for long-text processing and agent applications.

主要功能

• 百万上下文处理：原生支持1M Token超长文本理解与记忆，为官方服务标配。

  Million-token context handling: Natively supports understanding and memory of up to 1M tokens, standard for the official service.

• 混合注意力机制：CSA与HCA架构大幅降低长上下文计算与显存开销。

  Hybrid attention mechanism: CSA and HCA architectures significantly reduce computation and memory overhead for long contexts.

• 双模式推理：支持非思考与思考模式，后者可通过 reasoning_effort 参数调节强度。

  Dual-mode reasoning: Supports non-thinking and thinking modes; the latter can be tuned via the reasoning_effort parameter.

• 多领域专家融合：通过OPD蒸馏整合数学、代码、Agent等领域专家能力。

  Multi-domain expert fusion: Integrates expert capabilities in math, code, agents, etc., via OPD distillation.

• 经济高效选择：Flash版本用更低参数实现接近Pro的推理性能，API成本显著降低。

  Cost-efficient option: The Flash version achieves reasoning performance close to Pro with far fewer parameters, significantly reducing API costs.

技术原理

• CSA压缩稀疏注意力：将每m个token的KV压缩为1个条目，通过Lightning Indexer计算索引分数并执行Top-k稀疏选择，结合滑动窗口与Attention Sink机制保留局部依赖。

  CSA (Compressed Sparse Attention)压缩稀疏注意力，将每m个token的KV压缩为1个条目，通过稀疏选择降低长上下文计算量。: Compresses the KV of every m tokens into one entry, uses a Lightning Indexer to compute index scores and perform top‑k sparse selection, and retains local dependencies with sliding windows and Attention Sink.

• HCA重度压缩注意力：以更大压缩比m’将KV条目合并为单个条目，保持密集注意力而不采用稀疏选择，进一步降低计算量。

  HCA (Heavy Compression Attention)重度压缩注意力，以更大压缩比将KV条目合并为单个条目，保持密集注意力进一步降低计算量。: Merges KV entries into a single entry with an even larger compression ratio m’, maintaining dense attention without sparse selection to further reduce computation.

• mHC流形约束超连接：将残差映射矩阵通过Sinkhorn-Knopp算法投影到双随机矩阵流形，约束谱范数不超过1，增强深层信号传播稳定性。

  mHC (Manifold-Constrained Hyper-Connection): Projects the residual mapping matrix onto the doubly stochastic matrix manifold using the Sinkhorn-Knopp algorithm, constraining the spectral norm to ≤1 to enhance deep signal propagation stability.

• Muon优化器：模型采用混合Newton-Schulz迭代对梯度矩阵进行正交化，分快速收敛与精确稳定两个阶段，支持大规模MoE高效训练。

  Muon optimizer混合Newton-Schulz迭代对梯度矩阵正交化的优化器，支持大规模MoE高效训练。: Uses a hybrid Newton-Schulz iteration to orthogonalize the gradient matrix, divided into fast-convergence and precise-stability phases, enabling efficient training of large MoE models.

• FP4量化感知训练：对MoE专家权重和CSA索引器QK路径进行FP4量化，用FP8扩展动态范围实现无损反量化，降低内存与计算开销。

  FP4 quantization-aware training对MoE专家权重和CSA索引器QK路径进行FP4量化，用FP8扩展动态范围，降低内存与计算开销。: Quantizes MoE expert weights and the CSA indexer’s QK path to FP4, uses FP8 to extend dynamic range for lossless dequantization, reducing memory and computation overhead.

性能体现

知识能力

Knowledge Capabilities

Metric	DeepSeek-V4-Pro	Competitor Performance
SimpleQA-Verified	57.9%	Gemini-3.1-Pro: 75.6%
Chinese-SimpleQA	84.4%	K2.6: 75.9%, GLM-5.1: 75.0%
MMLU-Pro	87.5%	GPT-5.4: ≈87.5%, Gemini-3.1-Pro: >90%
GPQA Diamond	90.1%	GPT-5.4: ~90%, Gemini-3.1-Pro: >92%

推理与代码能力

Reasoning & Coding

Metric	DeepSeek-V4-Pro	Competitor Performance
HMMT 2026 Feb	95.2%	K2.6: ~92%, GLM-5.1: ~90%, GPT-5.4: ~96%
IMOAnswerBench	89.8%	GPT-5.4: ~91%, Opus-4.6: ~88%
Codeforces Rating	3206	GPT-5.4: 3168
Apex Shortlist	90.2%	GPT-5.4: 78.1%, Opus-4.6: 85.9%
LiveCodeBench	93.5%	All competitors: <90%

Agent能力

Agent Capabilities

Metric	DeepSeek-V4-Pro	Competitor Performance
SWE Verified	80.6%	Opus-4.6: 80.8%
SWE Pro	55.4%	–
SWE Multilingual	76.2%	–
Terminal Bench 2.0	67.9%	K2.6: 66.7%, GLM-5.1: 63.5%, Opus-4.6: 65.4%
MCPAtlas Public	73.6%	–
Toolathlon	51.8%	–

长上下文能力

Long-Context Capabilities

Metric	DeepSeek-V4-Pro	Competitor Performance
MRCR 1M	83.5%	Gemini-3.1-Pro: 76.3%
CorpusQA 1M	62.0%	Gemini-3.1-Pro: 53.8%

效率表现

Efficiency

Metric	DeepSeek-V4-Pro	DeepSeek-V4-Flash	Comparison (vs V3.2)
FLOPs per token (1M context)	27% of V3.2	10% of V3.2	–
Total KV cache (1M context)	10% of V3.2	7% of V3.2	–
Expert weight storage	FP4	–	Theoretical further 1/3 efficiency gain

如何使用DeepSeek-V4

• 网页端/App：访问 DeepSeek 官网或官方App，选择专家模式（Pro）或快速模式（Flash）。

  Web/App: Visit DeepSeek's official website or app, select Expert mode (Pro) or Fast mode (Flash).

• API调用：修改 model 参数为 deepseek-v4-pro 或 deepseek-v4-flash，base_url 保持不变。

  API call: Set the model parameter to deepseek-v4-pro or deepseek-v4-flash; keep base_url unchanged.

• 思考模式：复杂Agent场景建议启用思考模式并设置 reasoning_effort: max。

  Thinking mode: For complex agent scenarios, enable thinking mode and set reasoning_effort: max.

• 本地部署：通过Hugging Face或ModelScope下载开源权重自行部署。

  Local deployment: Download open-source weights from Hugging Face or ModelScope for self‑deployment.

关键信息和使用要求

• 版本规格：Pro版1.6T参数/49B激活，Flash版284B参数/13B激活，预训练数据分别为33T与32T。

  Model specs: Pro: 1.6T total / 49B active parameters; Flash: 284B total / 13B active parameters. Pre-training data: 33T and 32T tokens respectively.

• 上下文长度：两个版本均支持1M Token，旧接口 deepseek-chat 与 deepseek-reasoner 将于2026-07-24停用。

  Context length: Both versions support 1M tokens. Legacy interfaces deepseek-chat and deepseek-reasoner will be deprecated on 2026-07-24.

• API定价（每百万Token）：Pro输入缓存命中1元/未命中12元，输出24元；Flash输入缓存命中0.2元/未命中1元，输出2元。

  API pricing (per million tokens): Pro – input cached 1 RMB, uncached 12 RMB, output 24 RMB; Flash – input cached 0.2 RMB, uncached 1 RMB, output 2 RMB.

• 算力限制：Pro版当前服务吞吐有限，预计下半年昇腾950超节点批量上市后价格将大幅下调。

  Capacity constraints: The Pro version currently has limited throughput; prices are expected to drop significantly after the mass availability of Ascend 950 supernodes in the second half of the year.

核心优势

• 百万上下文普惠化：1M Token超长上下文成为官方服务标配，突破传统注意力机制的二次计算瓶颈，使长文本任务与测试时缩放真正可行。

  Democratizing million-token context: The 1M token ultra-long context becomes standard in the official service, breaking the quadratic computation bottleneck of traditional attention, making long-text tasks and test‑time scaling truly viable.

• 极致长上下文效率：通过CSA压缩稀疏注意力与HCA重度压缩注意力的混合架构，1M上下文下V4-Pro的单Token推理FLOPs仅为V3.2的27%，KV缓存仅10%，Flash版更是低至10%与7%。

  Extreme long-context efficiency: With the hybrid CSA+HCA architecture, V4-Pro's FLOPs per token at 1M context is only 27% of V3.2, and KV cache is only 10%; Flash version goes as low as 10% and 7% respectively.

• 开源模型性能新标杆：V4-Pro-Max在知识、推理、代码竞赛等评测中全面领先前代开源模型，Agent编码能力内部评测优于Claude Sonnet 4.5，交付质量接近Opus 4.6非思考模式。

  New open-source performance benchmark: V4-Pro-Max surpasses previous open-source models across knowledge, reasoning, and coding benchmarks; internal agent coding evaluations outperform Claude Sonnet 4.5 and approach Opus 4.6 non-thinking mode quality.

• 双版本灵活覆盖：Pro版（1.6T/49B）定位顶级性能，Flash版（284B/13B）以极小激活参数实现接近的推理能力，API价格低至Pro的1/12，普惠不同预算场景。

  Flexible dual‑version coverage: Pro (1.6T/49B) targets top-tier performance; Flash (284B/13B) achieves near‑Pro reasoning with minimal active parameters, API cost as low as 1/12 of Pro, serving various budget scenarios.

• Agent能力原生增强：针对Claude Code、OpenClaw等主流Agent框架专项优化，支持跨用户消息边界的连贯推理保留，在SWE、Terminal Bench等Agent评测中表现优异。

  Native agent enhancement: Specifically optimized for mainstream agent frameworks like Claude Code and OpenClaw; supports coherent reasoning retention across user message boundaries, excelling in agent benchmarks such as SWE and Terminal Bench.

项目地址

• HuggingFace模型库：https://huggingface.co/collections/deepseek-ai/deepseek-v4

  HuggingFace model collection: https://huggingface.co/collections/deepseek-ai/deepseek-v4

• 技术论文：https://huggingface.co/deepseek-ai/DeepSeek-V4-Pro/blob/main/DeepSeek_V4.pdf

  Technical paper: https://huggingface.co/deepseek-ai/DeepSeek-V4-Pro/blob/main/DeepSeek_V4.pdf

竞品对比

Competitor Comparison

Dimension	DeepSeek-V4-Pro	Claude Opus 4.6	Kimi K2.6
Model positioning	Open-source high-performance MoE	Closed-source top-tier general	Open-source agent intelligence
Open-source status	Fully open-source	Closed API	Open-source / open API
Total parameters	1.6T	Not disclosed	Not disclosed
Active parameters	49B	Not disclosed	Not disclosed
Context length	1M tokens	200K	1M tokens
Core architecture	CSA+HCA hybrid attention	Traditional Transformer	MoE + long context
MMLU-Pro	87.5	89.1	87.1
SimpleQA	57.9	46.2	36.9
Codeforces	3206	–	–
SWE Verified	80.6	80.8	80.2
Terminal Bench	67.9	65.4	66.7
MRCR 1M	83.5	92.9	–
API input price	12 RMB/M tokens	~150 RMB/M tokens	~60 RMB/M tokens
Long-context efficiency	KV cache = 10% of V3.2	Standard KV cache	Efficient but details not disclosed

应用场景

• 长文档分析：支持百万字级论文、报告、法律合同的全文理解与跨章节推理。

  Long-document analysis: Supports full-text understanding and cross‑chapter reasoning for millions-of-words papers, reports, and legal contracts.

• 智能体编码：在Claude Code、OpenClaw等框架中执行复杂代码生成、重构与调试任务。

  Agent coding: Executes complex code generation, refactoring, and debugging in frameworks like Claude Code and OpenClaw.

• 多轮工具调用：在Agent工作流中保留完整推理历史，支持跨用户消息边界的连贯思考。

  Multi-turn tool calling: Retains complete reasoning history in agent workflows, enabling coherent thinking across user message boundaries.

• 知识密集型问答：在世界知识评测中大幅领先开源模型，适用于教育、科研与专业咨询。

  Knowledge-intensive Q&A: Significantly outperforms open-source models in world knowledge benchmarks, suitable for education, research, and professional consulting.

• 白领办公任务：模型在中文写作、信息分析、文档生成与编辑等场景表现优异。

  Office automation tasks: Excels in Chinese writing, information analysis, document generation, and editing.

---

本文基于DeepSeek官方发布的技术资料整理，所有数据截至2026年4月。

This article is compiled from official DeepSeek technical materials; all data as of April 2026.

常见问题（FAQ）

DeepSeek-V4的上下文长度是多少？

DeepSeek-V4原生支持百万Token（1M）超长上下文，可处理大规模文本，适用于长文档分析和Agent应用。

DeepSeek-V4支持哪两种推理模式？

支持非思考模式和思考模式，思考模式可通过reasoning_effort参数调节推理强度，平衡性能与效率。

DeepSeek-V4 Pro和Flash版本有什么区别？

Pro版本参数更多，性能更高；Flash版本参数较少，但推理性能接近Pro，API成本显著降低，适合经济高效场景。