Realtime AI News
DeepSeek hit by global outage across chat and APIs as a mysterious Harness account appears
DeepSeek's chat service and APIs were broadly interrupted on the morning of Oct. 15, with the official status page showing the cause identified and a fix underway for DeepSeek chat, the V4 Pro API, and the V4.1 Flash API. A new Bilibili account named DeepSeekHarness appeared at the same time, and the report states there is no official evidence linking the outage to a Harness update or a new model.

DeepSeek's chat service and APIs went down for users around the world on the morning of Oct. 15. According to QbitAI, the company's official status page carried a red incident notice saying that parts of the web and API services were interrupted, that the cause had been identified, and that a fix was underway.
The affected services listed by the company include the DeepSeek chat service, the DeepSeek V4 Pro API, and the DeepSeek V4.1 Flash API — covering the consumer chat entry point and both developer-facing interfaces.
DeepSeek's incident record puts the first failure at 8:11 a.m. Beijing time on Oct. 15. The team initially said it was investigating and identifying the cause, then updated the status to "identified" and moved into repair. By the time the latest update was published, the disruption had lasted about an hour.
User reports varied: some sent messages and waited without a reply, some were stuck on a loading screen, and others hit prompts such as "server busy" or "too many messages, please try again later." Third-party monitoring platform StatusGator logged 86 related reports in the previous 24 hours, a number that was still climbing.
The overseas monitoring account Status Is DOWN also flagged the anomaly and asked whether other users were affected. A user in Luxembourg said the server did not respond, a user in Poland reported "no reply," and a user in New York summed it up as "so slow."
The speculation that arrived alongside the outage should be kept separate from the outage itself. QbitAI notes that every DeepSeek disruption brings a familiar round of comments asking whether a new model is coming, and this time an account named "DeepSeekHarness" appeared on Bilibili at the same moment — a freshly created page with little information, but already tens of thousands of followers.
DeepSeek Harness is an open-source agent runtime platform DeepSeek has already released, positioned around the idea that "everything is a plugin": the model does the thinking, while the harness calls tools, edits files, runs commands, handles code, and splits complex tasks across multiple sub-agents. The report reads the new Bilibili account as an effort to promote the agent product to Chinese users.
On whether the outage was preparation for a Harness update or a new model release, the report is explicit that there is no official evidence. The two signals do sit close together in time, but DeepSeek has not commented further; for developers, the practical things to watch are the status page recovery and actual API availability.
Why it matters
The direct effect is availability: roughly an hour of downtime across chat and two API tiers disrupts workflows for developers who depend on DeepSeek, with a third-party monitor logging dozens of reports. The second effect is informational, because the community habitually reads any outage as a model-release signal, a reading that should not be treated as fact without official evidence.
Nearby Updates
All10/15, 09:56
7天几百元开一家店!用千问办公电商版,立省启动资金5万元
7天几百元开一家店!用千问办公电商版,立省启动资金5万元. 给企业Agent打了个样
10/15, 09:55
1.38 亿笔链上支付之后,AI Agent 到底在买什么? 深潮TechFlow
1.38 亿笔链上支付之后,AI Agent 到底在买什么? 深潮TechFlow. 1.38 亿笔链上支付之后,AI Agent 到底在买什么? 深潮TechFlow
10/15, 10:13
Anthropic's 2021 internal research roadmap surfaces in copyright suit, sketching Claude Code years early
QbitAI reported on Oct. 15 that an internal Anthropic document marked "top secret" was exposed through a copyright lawsuit, and that it is a two-to-three page research roadmap written in July 2021. The file already named Claude Code, tool use, model self-judging, test-time compute, and trillion-parameter models with very large contexts, covering much of today's agent product language.
10/15, 10:45
OpenCode 2.0 Released, Letting Models Write Scripts to Orchestrate Tools
AI coding tool OpenCode 2.0 has officially launched, with its headline capability being that models can now write their own scripts to orchestrate tools, according to OSCHINA. That shifts coding agents from single tool calls toward scripted multi-step workflows, raising fresh questions about permissions and execution environments.