Back to Home

DeepSeek Just Capped Free Chat: How to Keep Working

On November 16, DeepSeek quietly started rationing its free chatbot. Users opened the app, typed an ordinary question, and got back a line that had never appeared before: "Due to heavy server load, your usage quota has reached the upper limit." No announcement, no pricing page, no date for when the caps lift. The company's stated reason is four words long: insufficient computing power.

If you use DeepSeek for anything longer than a quick lookup, this lands hard. A well-known lawyer who ran the model locally made headlines this month for squeezing it onto old GPUs, but most people do not have eight V100s in a cupboard. Here is what actually changed, what did not, and the routes around it that exist today.

What Changed on November 16

Three restrictions hit the consumer app and website at once, and none of them were formally published:

  • A daily conversation quota. Hit it and the app refuses new messages until the next day. Some users report being blocked after less than two hours of ordinary use.
  • Input capped at roughly 40,000 characters per prompt. That is comfortable for an essay and painfully tight for a 60-page PDF or a full repository dump.
  • Idle conversations expire. Threads with no activity for more than seven days can no longer be resumed, and a large number of users say they lost access after 72 hours instead.

The cut-offs are inconsistent between accounts, which is exactly what you would expect from a limit that was switched on before it was documented. DeepSeek has published no unified quota policy, and support has not said whether the caps will be relaxed or turned into paid tiers.

Developers on the paid API keep their own concurrency limits and pay-as-you-go billing. Those rules did not move. That gap is the most useful fact in this story: the same V4.1 Flash model that told a free user to come back tomorrow will answer all night if you bring a key.

The timing is hard to miss. On November 13, three days before the caps appeared, DeepSeek opened its full platform API, adding web search, file reading, page scraping and speech, and published an open-source demo client called chat-demo on GitHub. Squeeze the free window, widen the paid door. For anyone who needs sustained access, the API is no longer a workaround, it is the product DeepSeek is pointing you toward.

A Practical Playbook for Heavy Users

If the quota message has already interrupted real work, work through these in order. Most people only need the first two.

  1. Export before you lose it. Open any long thread you still care about and export the history today. Several users discovered the expiry rule only when the export button was the last thing still working.
  2. Move long projects onto an API key. Register on the platform, add a small balance, create a key, then run chat-demo locally or point any OpenAI-compatible client at DeepSeek's endpoint. The pay-as-you-go cost for reading and revising a document is measured in cents, not subscriptions.
  3. Chunk the input. Keep each request under about 30,000 characters and send a summary forward instead of the whole source. If you are feeding a paper or a codebase, summarize sections first and ask questions against the summary.
  4. Start fresh threads on purpose. Before a chat gets long, write a short brief that captures the state of the work, the decisions made and the open questions. Paste that into a new thread and keep the brief in a scratch file. You lose the model's memory, not your own.
  5. Move heavy runs off-peak. DeepSeek's August pricing already halves rates outside peak hours: V4 Pro drops to 13.5 yuan per million output tokens from 27, with input at 4.5 instead of 9. Flash runs 9 yuan at peak and 4.5 off-peak. If you are paying, the clock is a discount.

The Numbers Behind the Squeeze

This is not a company being stingy for its own sake. DeepSeek's V4.1-Flash-1017 model processed 59.2 trillion tokens on OpenRouter in the week ended November 15, first place worldwide. Free chat volume at that scale is a subsidy on every sentence, and the model that ranks first in calls also ranks first in inference bill.

The infrastructure signals line up. The status page has logged eight incidents since October affecting chat, V4 Pro or V4.1 Flash APIs. Expiring idle conversations is not arbitrary housekeeping either: dormant sessions hold KV cache resident in GPU memory, and dropping them frees VRAM for live traffic. A company short on accelerators, in a market where the best chips are export-controlled, has three places to take capacity from: training, paid API, or free chat.

There is also a strong hint about what comes next. Developers digging through the web front end found complete quota-management logic that syncs remaining usage every five minutes, plus a field named "Coding Plan" for a subscription tier that has never been announced. The most common comment on Chinese social platforms after the caps appeared was not outrage. It was that users would happily pay, if only there were somewhere to pay.

Treat free DeepSeek as a metered tool rather than a permanent workspace. Put quick questions in the app, put sustained or document-heavy work behind an API key, and keep your own notes rather than trusting a thread to survive the week. Until DeepSeek ships a paid consumer plan, the API is the only reliable substitute, and it is the one path the company has spent the past week making easier to walk.

The quota message is not a bug to wait out. It is a bill arriving in public, and the free tier is unlikely to grow back.

Comments

No comments yet. Be the first to share your thoughts!