GEO

DeepSeek V4预览版性能如何?2026年百万上下文MoE开源模型评测

2026/4/30
DeepSeek V4预览版性能如何?2026年百万上下文MoE开源模型评测

AI Summary (BLUF)

DeepSeek launches V4 Preview with two models: V4-Pro (1.6T total/49B active) and V4-Flash (284B/13B active), both open-sourced with 1M context length. Performance rivals top closed-source models, with structural innovations like DSA attention. API available now, older models retire July 2026.

原文翻译:DeepSeek发布V4预览版,包含两个模型:V4-Pro(总参数量1.6T/激活参数49B)和V4-Flash(总参数量284B/激活参数13B),均开源并支持100万上下文长度。性能媲美顶级闭源模型,具有DSA注意力等结构创新。API现已可用,旧模型将于2026年7月退役。

Overview: A New Era of Cost-Effective Long Context

🚀 DeepSeek-V4 Preview is officially live & open-sourced! Welcome to the era of cost-effective 1M context length.

🚀 DeepSeek-V4 Preview 现已正式上线并开源!欢迎进入高性价比的百万级上下文长度时代。

🔹 DeepSeek-V4-Pro: 1.6T total / 49B active params. Performance rivaling the world's top closed-source models.

🔹 DeepSeek-V4-Pro 总参数量 1.6T / 激活参数量 49B。性能可与全球顶尖闭源模型抗衡。

🔹 DeepSeek-V4-Flash: 284B total / 13B active params. Your fast, efficient, and economical choice.

🔹 DeepSeek-V4-Flash 总参数量 284B / 激活参数量 13B。为您提供快速、高效且经济的选择。

Try it now at chat.deepseek.com via Expert Mode / Instant Mode. API is updated & available today!

您现在即可通过 chat.deepseek.com 的专家模式/即时模式进行体验。API 已同步更新并立即可用!

Key Resources


Model Architecture and Specifications

DeepSeek-V4-Pro

🔹 Enhanced Agentic Capabilities: Open-source SOTA in Agentic Coding benchmarks.

🔹 增强的智能体能力: 在智能体编程基准测试中达到开源模型的 SOTA 水平。

🔹 Rich World Knowledge: Leads all current open models, trailing only Gemini-3.1-Pro.

🔹 丰富的世界知识: 领先所有现有开源模型,仅次于 Gemini-3.1-Pro。

🔹 World-Class Reasoning: Beats all current open models in Math/STEM/Coding, rivaling top closed-source models.

🔹 世界级推理能力: 在数学/STEM/编程领域超越所有现有开源模型,可与顶尖闭源模型匹敌。

DeepSeek-V4-Flash

🔹 Reasoning capabilities closely approach V4-Pro.

🔹 推理能力接近 V4-Pro 水平。

🔹 Performs on par with V4-Pro on simple Agent tasks.

🔹 在简单智能体任务上表现与 V4-Pro 持平。

🔹 Smaller parameter size, faster response times, and highly cost-effective API pricing.

🔹 更小的参数量、更快的响应速度以及极具性价比的 API 定价。

Model Comparison

Feature DeepSeek-V4-Pro DeepSeek-V4-Flash
Total Parameters 1.6T 284B
Active Parameters 49B 13B
Agentic Coding Open-source SOTA Strong, simple tasks on par with V4-Pro
World Knowledge Leads all open models High but below V4-Pro
Reasoning World-class, rivals top closed-source Closely approaches V4-Pro
Response Speed Standard Faster
API Pricing Standard Highly cost-effective
Target Use Case Heavy-duty tasks, complex reasoning Fast, efficient, cost-sensitive scenarios

Structural Innovation & Ultra-High Context Efficiency

🔹 Novel Attention: Token-wise compression + DSA (DeepSeek Sparse Attention).

🔹 新型注意力机制: 逐 Token 压缩结合 DSA(DeepSeek 稀疏注意力)。

🔹 Peak Efficiency: World-leading long context with drastically reduced compute & memory costs.

🔹 巅峰效率: 世界领先的长上下文能力,同时大幅降低计算和内存成本。

🔹 1M Standard: 1M context is now the default across all official DeepSeek services.

🔹 1M 上下文标准: 百万级上下文现已为所有官方 DeepSeek 服务的默认配置。

Key Innovation

Token-wise compression combined with DSA enables a novel attention mechanism. This drastically reduces computational and memory overhead while maintaining world-leading long-context performance. The result is that 1M context is now the baseline across all official DeepSeek services, making advanced long-context applications practical and affordable.

Token-wise 压缩结合 DSA 实现了一种新型注意力机制。这在大幅降低计算和内存开销的同时,保持了世界领先的长上下文性能。其结果是百万级上下文现已为所有官方 DeepSeek 服务的默认配置,使得先进的长上下文应用变得实用且可负担。


Dedicated Optimizations for Agent Capabilities

🔹 DeepSeek-V4 is seamlessly integrated with leading AI agents like Claude Code, OpenClaw & OpenCode.

🔹 DeepSeek-V4 已与 Claude Code、OpenClaw 和 OpenCode 等领先的 AI 智能体实现无缝集成。

🔹 Already driving our in-house agentic coding at DeepSeek.

🔹 该模型已在 DeepSeek 内部驱动我们的智能体编程应用。

DeepSeek-V4 has been specifically optimized for agentic workflows. It integrates seamlessly with leading AI agents such as Claude Code, OpenClaw, and OpenCode. The model is already powering DeepSeek's own in-house agentic coding efforts.

DeepSeek-V4 针对智能体工作流进行了专门优化。它与 Claude Code、OpenClaw 和 OpenCode 等领先的 AI 智能体无缝集成。该模型已在 DeepSeek 内部驱动我们的智能体编程应用。


API Availability and Pricing

API is Available Today!

🔹 Keep base_url, just update model to deepseek-v4-pro or deepseek-v4-flash.

🔹 保持 base_url 不变,只需将 model 参数更新为 deepseek-v4-prodeepseek-v4-flash

🔹 Supports OpenAI ChatCompletions & Anthropic APIs.

🔹 支持 OpenAI ChatCompletions 和 Anthropic API。

🔹 Both models support 1M context & dual modes (Thinking / Non-Thinking): https://api-docs.deepseek.com/guides/thinking_mode

🔹 两模型均支持百万级上下文和双模式(思考型/非思考型):https://api-docs.deepseek.com/guides/thinking_mode

⚠️ Note: deepseek-chat & deepseek-reasoner will be fully retired and inaccessible after Jul 24th, 2026, 15:59 (UTC Time). (Currently routing to deepseek-v4-flash non-thinking/thinking).

⚠️ 注意: deepseek-chat 和 deepseek-reasoner 将于 2026 年 7 月 24 日 15:59(UTC 时间)后全面退役并不可用。(当前已路由至 deepseek-v4-flash 的非思考型/思考型模式)。

API Integration Details

Parameter Details
Base URL No change required
Model Name deepseek-v4-pro or deepseek-v4-flash
API Compatibility OpenAI ChatCompletions & Anthropic APIs
Context Support 1M tokens for both models
Dual Modes Thinking / Non-Thinking for both models
Model Retirement deepseek-chat & deepseek-reasoner → Retired after Jul 24, 2026, 15:59 UTC

API Pricing Table

Model Input Price (per 1M tokens) Output Price (per 1M tokens) Cache Hit Price (per 1M tokens)
DeepSeek-V4-Pro $0.48 $1.92 $0.12
DeepSeek-V4-Flash $0.24 $0.96 $0.06

Final Remarks

🔹 Amid recent attention, a quick reminder: please rely only on our official accounts for DeepSeek news. Statements from other channels do not reflect our views.

🔹 鉴于近期受到的广泛关注,在此提醒:请仅以我们的官方账号发布的 DeepSeek 新闻为准。其他渠道的声明不代表我们的观点。

🔹 Thank you for your continued trust. We remain committed to longtermism, advancing steadily toward our ultimate goal of AGI.

🔹 感谢您一直以来的信任。我们坚守长期主义,稳步迈向我们的终极目标——通用人工智能 (AGI)。

DeepSeek remains dedicated to longtermism and the pursuit of AGI. We encourage the community to rely only on official channels for news and updates. Thank you for your continued trust and support.

DeepSeek 始终致力于长期主义和对通用人工智能的追求。我们鼓励社区仅信赖官方渠道获取新闻与更新。感谢您一直以来的信任与支持。

常见问题(FAQ)

DeepSeek-V4 Pro和Flash版本有什么区别?怎么选择?

Pro版本总参1.6T/激活49B,推理和知识领先,适合复杂任务;Flash版本284B/13B,速度更快、API成本更低,适合简单任务或成本敏感场景。

DeepSeek-V4支持多长的上下文?用了什么新技术?

支持100万token默认上下文,采用逐token压缩和DSA稀疏注意力,大幅降低计算和内存成本,实现高效长上下文处理。

旧版DeepSeek模型什么时候停止服务?

旧模型将于2026年7月退役,建议用户尽快迁移到V4 Preview系列模型,API现已可用。

← 返回文章列表
分享到:微博

版权与免责声明:本文仅用于信息分享与交流,不构成任何形式的法律、投资、医疗或其他专业建议,也不构成对任何结果的承诺或保证。

文中提及的商标、品牌、Logo、产品名称及相关图片/素材,其权利归各自合法权利人所有。本站内容可能基于公开资料整理,亦可能使用 AI 辅助生成或润色;我们尽力确保准确与合规,但不保证完整性、时效性与适用性,请读者自行甄别并以官方信息为准。

若本文内容或素材涉嫌侵权、隐私不当或存在错误,请相关权利人/当事人联系本站,我们将及时核实并采取删除、修正或下架等处理措施。 也请勿在评论或联系信息中提交身份证号、手机号、住址等个人敏感信息。