熟悉的剧情，又上演了。
8月4日上午，海外开源AI Coding Agent平台OpenCode率先发文：DeepSeek V4 Flash当前因前所未有的访问量出现容量不足问题，可能会遇到报错，正在紧急修复中。随后不少国内开发者也反馈，DeepSeek官方API上午几乎不可用，报错率飙升，响应延迟从正常的几百毫秒涨到几十秒。中午前后，DeepSeek官方确认问题已解决，服务恢复正常。
历史总是押韵的
为什么说熟悉？因为这一幕我们在不到一个月前刚见过——Kimi K3发布48小时后，月之暗面被迫暂停C端新用户订阅，原因一模一样：服务器负载接近上限，GPU集群扛不住调用量的指数级增长。再往前推，豆包、通义千问在流量高峰期都出过类似问题。
这不是技术事故，这是"幸福的烦恼"。一个模型因为能力强、价格低，用户用脚投票蜂拥而至，然后把服务打挂了。从这个角度说，能被用户打到宕机，其实是一种认可——说明你的模型确实有人用，而且是大量用。
但认可归认可，问题还是得解决。如果每次模型一好用就宕机，用户的信任是会被消耗的。尤其是To B场景，企业客户把你的API接入生产系统，结果你动不动就挂，谁敢用？C端用户可以等你恢复，企业客户的损失可是按分钟算的。
能力与容量的剪刀差
这件事暴露了国产大模型一个普遍问题：能力追得很快，容量建设没跟上。
看评测榜单，Qwen、DeepSeek、Kimi这些国产模型确实已经到了全球第一梯队，某些指标甚至超过了GPT和Claude。但评测是在受控环境下跑的，几百个并发请求就能测出最高分。真实世界是什么样的？是几万、几十万QPS的持续压力，是峰谷差巨大的流量波动，是各种复杂的边缘情况。这时候考验的就不是模型能力了，是整个推理基础设施的水平。
推理基础设施是个系统工程：GPU集群调度、跨云负载均衡、缓存策略、弹性扩缩容、熔断降级机制——每一环都得做好，才能在流量洪峰下不垮。OpenAI和Anthropic在这方面摸爬滚打了好几年，踩过无数坑才到今天的稳定性。国产模型厂商在模型能力上可以用一年走完别人三年的路，但基础设施的坑，该踩的一个都少不了。
另一个深层原因是成本压力。多储备GPU当然能抗流量，但GPU是要花钱的，而且很贵。如果按照峰值流量储备算力，平时大部分GPU都是闲置的，成本会非常难看。在价格战打得这么凶的今天（Qwen3.8-Max价格只有Opus5的四分之一），厂商其实没有太多余裕去备太多冗余算力。这是一个商业两难：不扩容会宕机，扩容了成本扛不住。
不过话说回来，这至少说明一件事：国产模型的需求是真实存在的，而且非常旺盛。如果没人用，想宕机都没机会。问题暴露了是好事，解决了就是进步。下次再发新模型的时候，希望各家能多备几张卡——毕竟用户等得起第一次，等不起第二次。
明天见。
A familiar story plays out again.
On the morning of August 4, the overseas open-source AI Coding Agent platform OpenCode was first to post: DeepSeek V4 Flash is experiencing capacity issues from unprecedented traffic, errors may occur, emergency fixes underway. Soon after, many domestic developers reported DeepSeek's official API was essentially unusable in the morning, error rates spiking, latency jumping from a normal few hundred milliseconds to tens of seconds. Around midday, DeepSeek officially confirmed the issue was resolved and service restored.
History Rhymes
Why familiar? Because we saw this exact scene less than a month ago — 48 hours after Kimi K3 launched, Moonshot AI was forced to pause new consumer subscriptions, for the exact same reason: server load approached limits, GPU clusters couldn't handle exponential traffic growth. Going further back, Doubao and Tongyi Qianwen all had similar issues during traffic peaks.
This isn't a technical failure — it's a "good problem to have". A model goes down because it's good, it's cheap, and users vote with their feet and flood in. From that perspective, getting hugged to death by users is actually a form of validation — it means your model is genuinely being used, and used heavily.
But validation doesn't solve the problem. If every time a model gets good it crashes, user trust gets eroded. Especially in B2B scenarios where enterprise customers wire your API into production systems — if you're going down left and right, who dares use you? Consumer users can wait for you to recover; enterprise customer losses are calculated by the minute.
The Capability-Capacity Scissors Gap
This incident exposes a common problem across domestic LLMs: capabilities are catching up fast, but capacity building hasn't kept pace.
Look at benchmarks — Qwen, DeepSeek, Kimi are genuinely in the global top tier, some metrics even beating GPT and Claude. But benchmarks run in controlled environments — a few hundred concurrent requests gets you top scores. What's the real world like? Tens to hundreds of thousands of sustained QPS, massive peak-valley swings, all kinds of edge cases. That's when it's not about model capability anymore — it's about the entire inference infrastructure.
Inference infrastructure is systems engineering: GPU cluster scheduling, cross-cloud load balancing, caching strategies, auto-scaling, circuit breaking and degradation — every ring has to be solid to survive traffic floods. OpenAI and Anthropic have been battle-testing this for years, stepping on countless mines to reach today's reliability. Domestic model vendors can cover three years of model capability progress in one, but infrastructure mines? Every single one still has to be stepped on.
Another underlying reason is cost pressure. Stockpiling more GPUs handles traffic, but GPUs cost money — lots of it. Reserving compute for peak traffic means most GPUs sit idle most of the time, which looks terrible on cost sheets. With the price war raging this hard (Qwen3.8-Max at a quarter of Opus 5 pricing), vendors don't have much margin for redundant compute. It's a commercial dilemma: don't scale and you crash; scale and costs become unsustainable.
But let's be fair — this at least proves one thing: demand for domestic models is real, and very strong. If nobody used you, you couldn't crash even if you tried. Exposing problems is good; solving them is progress. Next time a new model launches, hopefully everyone provisions a few more cards — users can forgive the first time, but not the second.
See you tomorrow.
能被用户打到宕机是一种认可，但每次好用就挂，企业客户谁敢把你接进生产系统？
—— Dawn Vision编辑部
Getting hugged to death by users is validation — but if you crash every time you get good, which enterprise customer dares wire you into production?
— The Dawn Vision Editorial Desk
DeepSeek宕机 · V4 Flash · 访问量过载 · 算力容量 · 国产模型 · Kimi K3教训 · 推理基础设施 · 能力容量剪刀差 · 价格战成本压力
DeepSeek outage · V4 Flash · traffic overload · compute capacity · domestic LLMs · Kimi K3 lesson · inference infrastructure · capability-capacity gap · price war cost pressure
Sources · 信源 Sources
本文基于 Dawn Vision 认知引擎处理的公开信息整理，素材来源：36氪、第一财经、OpenCode公告。
This article is based on public information processed by Dawn Vision. Sources: 36Kr, Yicai, OpenCode announcement.
