DeepSeek-V4预览版发布，百万上下文如何提升AI应用？

概述

DeepSeek-V4The next-generation model planned for release by DeepSeek in mid-February 2026, focusing on enhanced code generation and processing capabilities. 模型预览版于今日正式上线，并同步开源。该模型拥有百万字（1M tokens）超长上下文能力，在 Agent 能力模型自主执行任务、调用工具、进行代码生成等智能体相关能力。、世界知识和推理性能上均达到国内与开源领域的领先水平。模型根据参数量与成本分为两个版本：DeepSeek-V4The next-generation model planned for release by DeepSeek in mid-February 2026, focusing on enhanced code generation and processing capabilities.-Pro 和 DeepSeek-V4The next-generation model planned for release by DeepSeek in mid-February 2026, focusing on enhanced code generation and processing capabilities.-Flash。

The preview version of DeepSeek-V4The next-generation model planned for release by DeepSeek in mid-February 2026, focusing on enhanced code generation and processing capabilities. was officially released today, and the model weights have been open-sourced simultaneously. The model features a super-long context of 1 million tokens (1M), achieving state-of-the-art performance among domestic and open-source models in agent capabilities, world knowledge, and reasoning. Two versions are available based on parameter size and cost: DeepSeek-V4The next-generation model planned for release by DeepSeek in mid-February 2026, focusing on enhanced code generation and processing capabilities.-Pro and DeepSeek-V4The next-generation model planned for release by DeepSeek in mid-February 2026, focusing on enhanced code generation and processing capabilities.-Flash.

即日起，用户可通过官网 chat.deepseek.com 或官方 App 直接体验 1M 超长上下文对话。API 服务也已同步更新，指定 model_name 为 deepseek-v4-pro 或 deepseek-v4-flash 即可调用。

Starting today, users can experience the 1M long-context dialogue directly via the official website chat.deepseek.com or the official app. The API service has also been updated; simply set model_name to deepseek-v4-pro or deepseek-v4-flash to call the model.

---

开源链接与技术报告

The open-source model weights and technical report are available at the above links.

---

核心特性对比

DeepSeek-V4The next-generation model planned for release by DeepSeek in mid-February 2026, focusing on enhanced code generation and processing capabilities. 引入了全新的注意力机制，通过 token 维度压缩结合 DSA（DeepSeek Sparse Attention）稀疏注意力，在全球范围内实现了领先的长上下文能力，同时大幅降低计算和显存需求。以下是对两版本关键特性的多维对比：

The table above provides a multi-dimensional comparison of the two model versions, covering agent capabilities, world knowledge, reasoning performance, context length, supported modes, and API cost.

---

Agent 能力模型自主执行任务、调用工具、进行代码生成等智能体相关能力。深度分析

DeepSeek-V4The next-generation model planned for release by DeepSeek in mid-February 2026, focusing on enhanced code generation and processing capabilities. 针对主流 Agent 产品（如 Claude Code、OpenClaw、OpenCode、CodeBuddy）进行了专门的适配与优化，在代码任务、文档生成等方面均有显著提升。目前 DeepSeek-V4The next-generation model planned for release by DeepSeek in mid-February 2026, focusing on enhanced code generation and processing capabilities.-Pro 已被公司内部员工作为 Agentic Coding 模型使用，交付质量接近 Opus 4.6 非思考模式，但仍与 Opus 4.6 思考模式存在一定差距。

DeepSeek-V4The next-generation model planned for release by DeepSeek in mid-February 2026, focusing on enhanced code generation and processing capabilities. has been specifically adapted and optimized for mainstream agent products including Claude Code, OpenClaw, OpenCode, and CodeBuddy. It shows notable improvements in code tasks and document generation. Internally, DeepSeek-V4The next-generation model planned for release by DeepSeek in mid-February 2026, focusing on enhanced code generation and processing capabilities.-Pro is already used as the primary Agentic Coding model by company employees, delivering quality close to Opus 4.6 non-thinking mode, though still lagging behind Opus 4.6 thinking mode.

对于复杂的 Agent 场景，建议启用思考模式并将 reasoning_effort 设置为 max，以充分释放模型的深度推理潜力。

For complex agent scenarios, it is recommended to enable the thinking mode and set reasoning_effort to max to fully unleash the model’s deep reasoning potential.

---

技术亮点：百万上下文的实现

DeepSeek-V4The next-generation model planned for release by DeepSeek in mid-February 2026, focusing on enhanced code generation and processing capabilities. 的核心创新在于一种全新的注意力机制——在 token 维度进行压缩，并结合 DSA 稀疏注意力（DeepSeek Sparse Attention）。这一设计使得模型能够高效处理百万 token 的长序列，同时计算和显存需求相比传统方法大幅降低。

The core innovation of DeepSeek-V4The next-generation model planned for release by DeepSeek in mid-February 2026, focusing on enhanced code generation and processing capabilities. lies in a novel attention mechanism that compresses the token dimension and combines it with DSA (DeepSeek Sparse Attention). This design enables efficient processing of million-token sequences while significantly reducing compute and memory requirements compared to conventional methods.

官方宣布，从现在起，1M 上下文将成为 DeepSeek 所有官方服务的标配。

The official announcement states that from now on, 1M context will become the standard for all official DeepSeek services.

---

API 接口更新与迁移说明

现有 API 接口中，deepseek-chat 和 deepseek-reasoner 这两个模型名称将于 2026 年 7 月 24 日（三个月后）停止使用。在过渡期内，这两个名称将分别指向 deepseek-v4-flash 的非思考模式与思考模式。建议用户尽快迁移至新的 model_name：deepseek-v4-pro 或 deepseek-v4-flash。

The existing API model names deepseek-chat and deepseek-reasoner will be deprecated on July 24, 2026 (three months from now). During the transition period, these two names will map to the non-thinking and thinking modes of deepseek-v4-flash, respectively. Users are advised to migrate to the new model_name values (deepseek-v4-pro or deepseek-v4-flash) as soon as possible.

---

结语

DeepSeek-V4The next-generation model planned for release by DeepSeek in mid-February 2026, focusing on enhanced code generation and processing capabilities. 的发布标志着长上下文模型正式进入百万 token 时代，且开源版本在 Agent、知识和推理三大关键维度上均展现出顶尖实力。无论是追求极致性能的 Pro 版本，还是追求性价比的 Flash 版本，都为开发者和企业提供了强有力的选择。我们期待社区在开源模型基础上进行更多创新应用。

The release of DeepSeek-V4The next-generation model planned for release by DeepSeek in mid-February 2026, focusing on enhanced code generation and processing capabilities. marks the official entry of long-context models into the million-token era. The open-source versions demonstrate top-tier performance across the three key dimensions of agent, knowledge, and reasoning. Whether you choose the performance-oriented Pro version or the cost-effective Flash version, both offer powerful options for developers and enterprises. We look forward to more innovative applications built upon these open-source models by the community.

常见问题（FAQ）

DeepSeek-V4The next-generation model planned for release by DeepSeek in mid-February 2026, focusing on enhanced code generation and processing capabilities. 预览版有什么主要特点？

DeepSeek-V4The next-generation model planned for release by DeepSeek in mid-February 2026, focusing on enhanced code generation and processing capabilities. 预览版已正式上线并开源，拥有百万字超长上下文，Agent 能力模型自主执行任务、调用工具、进行代码生成等智能体相关能力。、世界知识和推理性能均领先，提供 Pro 和 Flash 两个版本供选择。

DeepSeek-V4The next-generation model planned for release by DeepSeek in mid-February 2026, focusing on enhanced code generation and processing capabilities.-Pro 和 DeepSeek-V4The next-generation model planned for release by DeepSeek in mid-February 2026, focusing on enhanced code generation and processing capabilities.-Flash 有什么区别？

Pro 版本在 Agent 能力模型自主执行任务、调用工具、进行代码生成等智能体相关能力。、世界知识和推理上更强，适合复杂任务；Flash 版本更经济快捷，简单任务与 Pro 相当。两者均支持 1M 上下文和思考模式，可通过官网或 API 使用。

如何获取 DeepSeek-V4The next-generation model planned for release by DeepSeek in mid-February 2026, focusing on enhanced code generation and processing capabilities. 的开源模型或调用 API？

开源模型可在 Hugging Face 和 ModelScope 下载；API 调用需设置 model_name 为 deepseek-v4The next-generation model planned for release by DeepSeek in mid-February 2026, focusing on enhanced code generation and processing capabilities.-pro 或 deepseek-v4The next-generation model planned for release by DeepSeek in mid-February 2026, focusing on enhanced code generation and processing capabilities.-flash。具体链接和技术报告详见文章。