⌘K
Change language Switch ThemeSign In
Narrow Mode
DeepSeek 网页版大升级！随后宕机 11 小时崩上热搜，新模型真的来了
量 量子位 @梦晨
One Sentence Summary
DeepSeek's web interface experienced a prolonged outage, likely due to a silent model upgrade. Users reported significant improvements in self-identification, knowledge recency, and code generation capabilities.
Summary
This article covers the recent prolonged outage of the DeepSeek web interface, interpreting it as a precursor to a major model upgrade. Before the service interruption, multiple users observed a leap in model capabilities, including more complex SVG composition and higher-quality frontend code generation. Key evidence suggests the new model version explicitly identifies itself as DeepSeek-V3, with a knowledge cutoff date updated to January 2026. Although service has been partially restored, instability persists, including bugs in the 'Deep Think' mode and suspected version rollbacks. Combined with DeepSeek's recent large-scale recruitment for Agent-related roles, the industry generally believes they are preparing for the release of a next-generation model.
Main Points
* 1. The DeepSeek web outage is viewed as a signal of a silent model upgrade.DeepSeek has a consistent habit of upgrading models without prior announcements. This 11-hour service interruption was accompanied by significant enhancements in model capabilities observed by users just before the crash. * 2. The model exhibits new version characteristics in self-identification and knowledge recency.The upgraded version consistently identifies itself as DeepSeek-V3 and demonstrates knowledge of international events from early 2026 without web access, suggesting a substantial update to its pre-training data. * 3. Practical application capabilities (drawing and coding) have seen visible improvements.User testing shows that the model is superior to previous versions in handling complex SVG compositions and generating logical, aesthetically pleasing frontend code in a single pass. * 4. Service remains unstable after restoration, with functional bugs.Currently, there is a severe bug in the 'Deep Think' mode where it 'thinks but fails to output an answer,' and some users have reported that the model version appears to have temporarily rolled back to the old version.
Metadata
AI Score
86
Website qbitai.com
Published At Today
Length 691 words (about 3 min)
Sign in to use highlight and note-taking features for a better reading experience. Sign in now
> 梦晨 发自 凹非寺 > > 量子位 | 公众号 QbitAI DeepSeek崩上热搜！
宕机持续超过8小时，写论文的、角色扮演的和心情不好找AI吐槽的人也都崩溃了。
But！这不是一次普通的服务中断，反而被解读为模型升级的前兆。
就在宕机前，很多用户报告DeepSeek网页版它变了，模型能力大幅提升。
比如经典的用SVG画鹈鹕骑自行车任务，3月29日的DeepSeek，构图和色彩都明显比一周前强。
而且DeepSeek这家公司，不预告、不公告，瞧瞧静默升级模型，已经不是第一次了。
模型到底有没有真的升级，答案就藏在对话记录里。 首先是身份标识更新，3月29日的版本会稳定的介绍自己为DeepSeek-V3模型。
一周前，同样的问题，它的自我介绍还是模糊的“我是DeepSeek，一款纯文字AI助手……“，对自己的版本号只字不提。
如果说这种变化也可能是偶然，那么知识截止日期的变化更有说服力。
在不开启联网搜搜的情况下，现在它知道25年的美国选举结果。但不知道26年2月的大事，有人据此推测新版本的知识截止日期可能是26年1月。
最后，3月29日DeepSeek一次性写代码生成前端页面的效果也有大幅提高。
所以到底是微调后的V3，还是直接升级了V4？
DeepSeek官方一个字都没说。
目前DeepSeek网页版已恢复服务。
不过还是有点小问题，开启深度思考模式的话，模型在思考结束后就会停止输出，并不会把答案输出在正文里。
不开深度思考模式的话，从模型自我介绍看似乎已恢复成老版本。
总之，结合DeepSeek上周一口气开启17个岗位，大招Agent方向人才来看：
在龙虾时代沉默了好久的DeepSeek，似乎在憋个大的。
参考链接：
[1]https://x.com/marmaduke091/status/2038211313566965823?s=20
[2]https://x.com/AiBattle_/status/2038206244016537615?s=20
量 量子位 @梦晨
One Sentence Summary
DeepSeek's web interface experienced a prolonged outage, likely due to a silent model upgrade. Users reported significant improvements in self-identification, knowledge recency, and code generation capabilities.
Summary
This article covers the recent prolonged outage of the DeepSeek web interface, interpreting it as a precursor to a major model upgrade. Before the service interruption, multiple users observed a leap in model capabilities, including more complex SVG composition and higher-quality frontend code generation. Key evidence suggests the new model version explicitly identifies itself as DeepSeek-V3, with a knowledge cutoff date updated to January 2026. Although service has been partially restored, instability persists, including bugs in the 'Deep Think' mode and suspected version rollbacks. Combined with DeepSeek's recent large-scale recruitment for Agent-related roles, the industry generally believes they are preparing for the release of a next-generation model.
Main Points
* 1. The DeepSeek web outage is viewed as a signal of a silent model upgrade.
DeepSeek has a consistent habit of upgrading models without prior announcements. This 11-hour service interruption was accompanied by significant enhancements in model capabilities observed by users just before the crash.
* 2. The model exhibits new version characteristics in self-identification and knowledge recency.
The upgraded version consistently identifies itself as DeepSeek-V3 and demonstrates knowledge of international events from early 2026 without web access, suggesting a substantial update to its pre-training data.
* 3. Practical application capabilities (drawing and coding) have seen visible improvements.
User testing shows that the model is superior to previous versions in handling complex SVG compositions and generating logical, aesthetically pleasing frontend code in a single pass.
* 4. Service remains unstable after restoration, with functional bugs.
Currently, there is a severe bug in the 'Deep Think' mode where it 'thinks but fails to output an answer,' and some users have reported that the model version appears to have temporarily rolled back to the old version.
Key Quotes
* This is not just an ordinary service interruption; it is being interpreted as a precursor to a model upgrade. * DeepSeek, as a company, has a history of silently upgrading models without any prior notice or announcement. * The knowledge cutoff date for the new version appears to be January 2026. * After a long period of silence, DeepSeek seems to be cooking up something big.
AI Score
86
Website qbitai.com
Published At Today
Length 691 words (about 3 min)
Tags
DeepSeek
Model Upgrade
DeepSeek-V3
LLM Outage
Code Generation
Related Articles
* 150: Year-End AI Review: From Models to Applications, Technology to Business Wars, Grasping the Thread of Meaning in the Torrent * Google's New Paper Crashes Memory Stocks! 6x KV Cache Compression, Netizens: Silicon Valley Becomes Reality * Vol.86 Same Generation Technology, Two Systems: A 181-Page PPT Record of the AI Industry in 2025 * #422. OpenAI Inside Secrets: How to Build a 'One-Person Billion-Dollar Company' with AI? A Conversation with Sherwin Wu, Head of API * AI Starts to "Take Action", Alibaba's Qwen Leads the World * Yao Shunyu Lectures Face-to-Face with Tang Jie, Yang Zhilin, and Lin Junyang! Four Schema Heroes Debate Heroes at Zhongguancun * Podcast Conversation and Interview: Sebastian Raschka and Nathan Lambert Deeply Interpret the State of AI on Lex Fridman Podcast: US-China Competition, Model Comparisons, Scaling Laws, and AGI Timelines. * Practices and Reflections on Vibe Coding in Code Generation and Collaboration * Zhipu GLM-5 Technical Deep Dive: Full Compatibility with Domestic Chips Like Huawei Sparks Global Discussion and asynchronous reinforcement learning infrastructure, while completing full-stack adaptation for domestic chips.") * GPT-5.4 Released: OpenAI's First Unified Model, Truly Native HomeArticlesPodcastsVideosTweets