← Back to stories Contemporary computer on support between telecommunication racks and cabinets in modern data center
Photo by Brett Sayles on Pexels
IT之家 2026-03-30

DeepSeek Quietly Upgraded Again, Nightlong Downtime Takes Users Back to Square One

Outage and frustration

DeepSeek (DeepSeek) quietly rolled out another model upgrade on March 29, but a nightlong outage wiped away the early gains and left many users frustrated. The web service was down for more than eight hours; it has been reported that some users experienced up to 11 hours of downtime. Students writing papers, role‑players and people looking for an AI to vent to all found the service unusable during the outage.

What changed — and what didn’t

Reporters and users noticed immediate improvements after the upgrade: image composition and color in SVG tasks—think the classic “pelican riding a bicycle” prompt—looked noticeably better than a week earlier, and front‑end code generation for web pages showed a marked improvement. The model now self‑identifies as DeepSeek‑V3. It reportedly knows the result of the U.S. 2025 elections when offline, but it does not seem aware of events from February 2026, so some observers speculate its knowledge cutoff may be around January 2026. A lingering bug remains: in “deep thinking” mode the model can stop producing output after completing internal reasoning and not surface the answer in the response body.

Strategy and wider context

The silent, unannounced upgrade is not new for the company. It has been reported that DeepSeek previously pushed quiet updates, and last week the firm posted 17 new job openings targeting Agent‑direction talent — a hint that it is gearing up for broader agent‑style features. Why the secrecy? In China’s fast‑moving AI ecosystem, rapid iteration and stealth pushes are common as firms race to catch up with global rivals and to mitigate exposure to geopolitical risks such as export controls on advanced chips. Will this iteration be the start of a more transparent upgrade cadence, or another brief blip that users must weather? For now, the best answer is experimentation: users and third‑party testers will decide whether DeepSeek‑V3 is actually a step forward.

AISmartphones
View original source →