本文来源:https://www.limbo.news/news/ingested-v54cu
若有版权问题,请来信告知:13951431913#139.com 或在右方提交评论。
DeepSeek发布V4模型,性能提升显著
DeepSeek官方宣布推出V4模型,该版本在性能上实现显著提升,具体细节尚未公布。
DeepSeek作为AI领域的新兴力量,其V4模型的发布可能对行业竞争格局产生影响。对于关注AI技术发展的读者,了解这一进展有助于把握技术趋势。
以下为抓取到的原文内容译文,已统一为站内阅读格式。
DeepSeek-V4 预览版:迈入百万上下文普惠时代
今天,我们全新系列模型 DeepSeek-V4 的预览版本正式上线并同步开源。
DeepSeek-V4 拥有百万字超长上下文,在 Agent 能力、世界知识和推理性能上均实现国内与开源领域的领先。模型按大小分为两个版本:
DeepSeek-V4-Pro:性能比肩顶级闭源模型
Agent 能力大幅提高: 相比前代模型,DeepSeek-V4-Pro 的 Agent 能力显著增强。在 Agentic Coding 评测中,V4-Pro 已达到当前开源模型最佳水平,并在其他 Agent 相关评测中同样表现优异。目前 DeepSeek-V4 已成为公司内部员工使用的 Agentic Coding 模型,据评测反馈使用体验优于 Sonnet 4.5,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。
丰富的世界知识: DeepSeek-V4-Pro 在世界知识测评中,大幅领先其他开源模型,仅稍逊于顶尖闭源模型 Gemini-Pro-3.1。
世界顶级推理性能: 在数学、STEM、竞赛型代码的测评中,DeepSeek-V4-Pro 超越当前所有已公开评测的开源模型,取得了比肩世界顶级闭源模型的优异成绩。
DeepSeek-V4-Flash:更快捷高效的经济之选
相比 DeepSeek-V4-Pro,DeepSeek-V4-Flash 在世界知识储备方面稍逊一筹,但展现出了接近的推理能力。而由于模型参数和激活更小,相较之下 V4-Flash 能够提供更加快捷、经济的 API 服务。
在 Agent 测评中,DeepSeek-V4-Flash 在简单任务上与 DeepSeek-V4-Pro 旗鼓相当,但在高难度任务上仍有差距。
DeepSeek-V4 开创了一种全新的注意力机制,在 token 维度进行压缩,结合 DSA 稀疏注意力(DeepSeek Sparse Attention),实现了全球领先的长上下文能力,并且相比于传统方法大幅降低了对计算和显存的需求。从现在开始,1M(一百万)上下文将是 DeepSeek 所有官方服务的标配。
DeepSeek-V4 针对 Claude Code、OpenClaw、OpenCode、CodeBuddy 等主流的 Agent 产品进行了适配和优化,在代码任务、文档生成任务等方面表现均有提升。下图为 V4-Pro 在某 Agent 框架下生成的 PPT 内页示例:
目前,DeepSeek API 已同步上线 V4-Pro 与 V4-Flash,支持 OpenAI ChatCompletions 接口与 Anthropic 接口。访问新模型时,base_url 不变,model 参数需要改为 deepseek-v4-pro 或 deepseek-v4-flash。
V4-Pro 与 V4-Flash 最大上下文长度为 1M,均同时支持非思考模式与思考模式,其中思考模式支持 reasoning_effort 参数设置思考强度(high/max)。对于复杂的 Agent 场景建议使用思考模式,并设置强度为 max。模型调用与参数调整方法请参考 思考模式指南 - API 文档 。
感谢每一位用户的信任与支持,大家的肯定、建议和期许,是我们不竭探索、持续进步的动力,也让我们始终坚守初心,专注于不懈的创新。
我们将始终秉持长期主义的原则理念,在尝试与思考中踏实前行,努力向实现 AGI 的目标不断靠近。
本文来源:https://www.limbo.news/news/ingested-v54cu
若有版权问题,请来信告知:13951431913#139.com 或在右方提交评论。
DeepSeek releases V4 model with significant performance improvements
DeepSeek officially announced the launch of the V4 model, which delivers significant performance improvements, though specific details have not yet been disclosed.
As an emerging force in the AI field, DeepSeek's release of the V4 model could have an impact on the industry's competitive landscape. For readers following AI technology developments, understanding this progress helps in grasping technical trends.
The following is a translation of the original captured content, standardized into the site's reading format.
DeepSeek-V4 Preview: entering the era of universal million-token context
Today, the preview version of our brand-new model series, DeepSeek-V4, is officially online and open-sourced at the same time.
DeepSeek-V4 features an ultra-long context of one million tokens, leading domestically and in the open-source domain in terms of agent capabilities, world knowledge, and reasoning performance. The model comes in two versions by size:
DeepSeek-V4-Pro: performance on par with top closed-source models
Significantly enhanced agent capabilities: Compared with the previous generation, DeepSeek-V4-Pro's agent capabilities are markedly stronger. In Agentic Coding evaluations, V4-Pro has reached the best level among current open-source models and performs excellently in other agent-related benchmarks as well. At present, DeepSeek-V4 has become the Agentic Coding model used internally by company employees; feedback indicates the experience is better than Sonnet 4.5, with delivery quality close to Opus 4.6's non-thinking mode, though still lagging somewhat behind Opus 4.6's thinking mode.
Rich world knowledge: In world knowledge evaluations, DeepSeek-V4-Pro significantly outpaces other open-source models, falling only slightly short of the top closed-source model Gemini-Pro-3.1.
World-class reasoning performance: In evaluations of mathematics, STEM, and competitive programming, DeepSeek-V4-Pro surpasses all currently publicly evaluated open-source models, achieving excellent results on par with the world's top closed-source models.
DeepSeek-V4-Flash: a faster, more efficient, and economical choice
Compared with DeepSeek-V4-Pro, DeepSeek-V4-Flash is slightly less capable in terms of world knowledge reserves, but demonstrates nearly comparable reasoning ability. Thanks to smaller model parameters and activations, V4-Flash can provide faster and more economical API services by comparison.
In agent evaluations, DeepSeek-V4-Flash is on par with DeepSeek-V4-Pro on simple tasks, but still lags on high-difficulty tasks.
DeepSeek-V4 pioneers a brand-new attention mechanism that compresses along the token dimension; combined with DSA Sparse Attention (DeepSeek Sparse Attention), it achieves world-leading long-context capabilities while significantly reducing compute and VRAM requirements compared with traditional methods. From now on, 1M (one million) context will be standard across all DeepSeek official services.
DeepSeek-V4 has been adapted and optimized for mainstream agent products such as Claude Code, OpenClaw, OpenCode, and CodeBuddy, with improved performance in code tasks, document generation tasks, and more. The figure below shows an example of a PPT interior slide generated by V4-Pro under a certain agent framework:
Currently, the DeepSeek API has launched both V4-Pro and V4-Flash simultaneously, supporting the OpenAI ChatCompletions interface and the Anthropic interface. When accessing the new models, the base_url remains unchanged, while the model parameter needs to be changed to deepseek-v4-pro or deepseek-v4-flash.
Both V4-Pro and V4-Flash support a maximum context length of 1M, and both support non-thinking and thinking modes; in thinking mode, the reasoning_effort parameter can be used to set thinking intensity (high/max). For complex agent scenarios, thinking mode is recommended with intensity set to max. For model invocation and parameter adjustment, please refer to the Thinking Mode Guide - API Documentation.
Thank you to every user for their trust and support. Your affirmation, suggestions, and expectations are the driving force behind our endless exploration and continuous progress, and they keep us true to our original mission, focused on relentless innovation.
We will always uphold the principle of long-termism, moving steadily forward through experimentation and reflection, and striving ever closer to the goal of achieving AGI.
