DeepSeek Down: Partial Outage Confirmed, the 'Mystery Harness Account' Is Not a Mystery

Verdict: DeepSeek is having its worst day in months: official status page confirms a partial outage of both the web chat and the V4.1 Flash API, cause identified and fix in progress, roughly two hours in at time of writing. If your pipeline runs on deepseek-flash, this is a real production event — the same API that already shows the lowest 90-day uptime on DeepSeek's own board (99.55%). Meanwhile, a "mystery account" fuss in Chinese tech media turned out to be DeepSeek's official Harness repository on GitHub — a 249,800-star open-source project tagged "Everything is a Plugin". Two unrelated stories, one useful takeaway: your DeepSeek dependency deserves an outage plan.

What the status page actually says

Verified directly on status.deepseek.com at the time of writing: banner reads "We're currently experiencing issues", incident titled "DeepSeek Web/API Partially Unavailable", status "Identified · Ongoing for 2 hours · Affects Chat Service, DeepSeek V4.1 Flash API Service, and 1 more", with the note "The cause has been identified and a fix is being implemented." A Reuters report from March 2026 documented a then-record 7-hour outage; today's incident started well inside that envelope, but we'll note it here if it stretches.

The 90-day uptime board tells you who got hurt worst: V4.1 Flash API at 99.55%, the lowest of every listed component — chat sits at 99.53% only if you count its three sub-services separately. For a model marketed at high-throughput production workloads, four nines-adjacent availability is the real price you pay on top of the token price. That 0.45% of downtime lands on somebody's batch job every month.

The "mystery account" is not a mystery

Chinese tech media spent the day pairing the outage with a "mysterious Harness account appearing at the same time." We checked the thing directly: github.com/deepseek-ai/deepseek-harness is DeepSeek's official org repository, created August 13, 2026, described as "DeepSeek Harness: Everything is a Plugin," currently sitting at roughly 250k stars, with active user discussions. There is no shadow product leak here — the repo has been public for two months, and its discussion board even carries a running joke about "servers overloaded," which is precisely what today's outage made literal.

We've covered Harness before as the plugin architecture behind DeepSeek's agent ecosystem. If today's outage pushes DeepSeek to ship more of its infrastructure as self-hostable plugins, that's the one silver lining worth watching: less dependence on their uptime, more of their tooling on your machines.

What to do while it's down

Practical order of operations: check status.deepseek.com before burning retries — partial outages often keep some endpoints alive while others time out. If your workload can wait, batch jobs are better queued than canceled; fix-in-progress incidents typically resolve faster than unexplained ones. If you can't wait, this is exactly the case for a fallback route — most agent stacks let you point the same prompt at a second provider when the primary errors out. And if you keep hitting failures after the banner clears, clear your connection or check our guide to DeepSeek's "server busy" errors — residual congestion after an outage is common and looks identical to the outage itself.

We run this site's own pipeline on DeepSeek's API, so today was spent practicing what this article preaches: status page first, retries second. The math of cheap tokens doesn't care which model we like — it just occasionally reminds you why the discount comes with a footnote.

Sources: DeepSeek official status page (2026-10-15); GitHub API, deepseek-ai/deepseek-harness (2026-10-15); Reuters, 2026-03-29. No affiliate links.