DeepSeek's free ride just got narrower. On November 16, users of the Chinese lab's consumer app and web chat began reporting a set of new ceilings, all justified by the same phrase: insufficient compute. Long conversations that have sat inactive for more than about a week can no longer be continued; single message inputs are now capped at roughly 40,000 characters; and a daily usage quota cuts off heavy users, sometimes in under two hours, with a notice that reads "server busy, usage limit reached."
The changes were first flagged by Chinese media including CNR and Guangming Daily after the topic trended on Chinese social platforms. What makes the rollout controversial is not the restriction itself but its opacity: DeepSeek has published no unified rule set, no explanation of why some users see a seven-day cutoff while others report seventy-two hours, no timeline for when limits might lift, and no paid tier that would let heavy users buy their way out.
The company's own status page explains why the squeeze is happening. It records at least eight service incidents since the start of October, including a severe outage on October 15 that kept both the consumer chat service and the V4 Pro and V4.1 Flash APIs down for close to five hours. Demand, meanwhile, has exploded: OpenRouter's weekly rankings showed DeepSeek V4.1 Flash processing 33.6 trillion tokens in the week of October 19-25, up 801 percent week over week and the largest volume of any model on the platform.
Analysts quoted in Chinese coverage point to the arithmetic of long conversations: every new token attends over the whole context, and inference frameworks cache those intermediate results in GPU memory. Idle threads that users may never return to keep occupying scarce VRAM, so cutting off week-old conversations and trimming input sizes is effectively a cache-reclamation policy aimed at keeping active users served.
Notably, the developer API is unaffected. API access has its own quotas and per-token billing, so the caps apply only to the free consumer surfaces. That is consistent with how DeepSeek has managed scarcity before: it paused API top-ups in February 2025 citing server strain, and in August 2026 moved the API to peak and off-peak pricing with off-peak calls at half price, using rules rather than revenue to flatten demand.
The missing piece is the consumer monetization that every other major lab has already built. Users flooded the app's review sections asking for exactly that, saying they would happily pay for stable access to long-running roleplay, research and document-editing sessions that the new limits break. DeepSeek has so far declined, keeping its consumer product free while quietly converting "free and unlimited" into "free within a budget you did not agree to."
For the industry, the episode is a data point on what runaway open-model demand looks like from the inside: even a lab that trains efficiently and gives its best models away cannot provision inference fast enough when usage grows eightfold in a week. The quota wall is the honest version of that constraint, and DeepSeek's silence about it may cost it more goodwill than the limits themselves.
Comments (0)
Log in to join the discussion
Log InNo comments yet