DeepSeek V4 发布了。这是最佳的开源编码模型。以下是详细介绍。
2 分•作者: Alisaqqt•3 个月前
两种模型:Flash(总参数量 2840 亿,激活参数量 130 亿)和 Pro(总参数量 1.6 万亿,激活参数量 490 亿)。两者都支持 100 万 token 上下文。
V4-Pro 是他们的旗舰模型。在 Agent 编码任务上击败了 Claude Opus 4.6 Max(他们自己说的)。特别提到在编码方面优于 Sonnet 4.5,并且在通用基准测试中与 Opus 4.6 相当。在世界知识和 STEM 方面,他们表示它领先于 Gemini-Pro-3.1。
V4-Flash 是一个被低估的选择。它比 Pro 更快、更便宜,但它比 Pro 具有更好的长上下文效率。
原文:
Agent 能力大幅提升:V4-Pro 在开源模型中,Agentic 编码基准测试中达到 SOTA 水平。实际上,用户报告说它感觉比 Sonnet 4.5 更好,并且输出质量接近 Opus 4.6 的非思考模式——尽管与启用思考的 Opus 4.6 相比,仍然存在差距。
世界知识:V4-Pro 在知识基准测试中以显著优势领先于所有开源模型,仅次于闭源前沿模型中的 Gemini-Pro-3.1。
顶级推理能力:在数学、STEM 和竞争性编码方面,V4-Pro 击败了所有已公开基准测试的开源模型,并且与世界上最好的闭源模型展开激烈竞争。
100 万上下文才是真正的亮点。完全重新设计了注意力机制——结合了称为 DSA(深度稀疏注意力)的技术来处理这种规模,而不会导致计算量爆炸。与 V3.2 相比,V4 的推理成本随着 token 数量的增加保持不变,而 V3.2 的成本会急剧上升。架构的改进使得它真正可用,而不仅仅是一个规格数字。
Agent 能力得到了专门的升级。专门针对 Claude Code、OpenClaw、OpenCode 和 CodeBuddy 进行了训练。V4-Pro 现在是任何 agentic / 编码工作流程的推荐模型。Flash 明确不推荐用于最复杂的 agent 任务。
API 已上线。定价:
DeepSeek-V4-Flash:每百万输入/输出 token 费用分别为 0.14 美元 / 0.28 美元
DeepSeek-V4-Pro:每百万输入/输出 token 费用分别为 1.74 美元 / 3.48 美元
Reasoning_effort 参数允许您为每次调用设置思考强度(低/高/最大)。特别推荐将“最大”用于 agent 任务。
该模型将在 Atlas Cloud 上发布。开发者可以获得 API 访问权限。
查看原文
Two models: Flash (284B total, 13B active) and Pro (1.6T total, 49B active). both hit 1M token context.<p>V4-Pro is their flagship. Beats Claude Opus 4.6 Max on Agent coding tasks (their words). specifically calls out being better than Sonnet 4.5 on coding, and competitive with Opus 4.6 on general benchmarks. on world knowledge and STEM, they say it's ahead of Gemini-Pro-3.1.<p>V4-Flash is the sleeper pick. Faster and cheaper than Pro, but it has better long-context efficiency than Pro does.<p>Original Text:
Agent capabilities massively improved: V4-Pro hits SOTA on Agentic Coding benchmarks among open-source models. In practice, users report it feels better than Sonnet 4.5, and output quality is close to Opus 4.6 non-thinking mode — though there's still a gap vs Opus 4.6 with thinking enabled.<p>World knowledge: V4-Pro leads all open-source models by a significant margin on knowledge benchmarks, sitting just behind Gemini-Pro-3.1 among closed-source frontier models.<p>Top-tier reasoning: On math, STEM, and competitive coding, V4-Pro beats every open-source model that's been publicly benchmarked and is trading blows with the best closed-source models in the world.
the 1M context is the real headline. Redesigned attention entirely — combines something called DSA (Deeply Sparse Attention) to handle the scale without blowing up compute. V4 inference cost stays flat as tokens scale up vs V3.2 which shoots up. the architecture improvement is what makes this actually usable, not just a spec number.<p>Agent capabilities got a dedicated upgrade. Trained specifically against Claude Code, OpenClaw, OpenCode, and CodeBuddy. V4-Pro is now the recommended model for any agentic / coding workflow. Flash is explicitly not recommended for the most complex agent tasks.<p>API is live. Pricing:<p>DeepSeek-V4-Flash: $0.14 / $0.28 per M input/output tokens<p>DeepSeek-V4-Pro: $1.74 / $3.48 per M input/output tokens<p>Reasoning_effort parameter lets you set thinking intensity (low/high/max) per call. "max" is recommended for agent tasks specifically.<p>The model will launch on Atlas Cloud. Developers can get API access.