Sometime between November 15 and 16, users of DeepSeek's free App and web chat started hitting walls that had not been there the week before. The company has not published a rule sheet. What follows is assembled from user reports and press checks, and it should be read that way.
The three limits
Long threads now expire. A conversation idle for more than seven days cannot be continued; some users report being cut off at 72 hours instead, with a message citing insufficient compute and suggesting the thread be exported elsewhere.
Single inputs are capped at roughly 40,000 characters. Paste more and the App refuses with a "content too long, please shorten" prompt rather than truncating silently.
Daily usage is metered. When the allowance runs out the App returns a generic "server busy" message. Reports vary on how fast that happens; some users hit it inside two hours of ordinary use.
DeepSeek's status page showed normal operation through the change, which is the tell. This was not an outage. It was a decision.
Why it happened
The number circulating is call volume. V4.1-Flash has been processing more than 33 trillion tokens a week, roughly four times the volume of three months earlier, and by several accounts DeepSeek now sits first in global token throughput. Capacity did not grow fourfold in a quarter.
The technical pressure is concentrated in one place: the KV cache. Long conversations are expensive not because the model is slow but because every additional turn has to keep the attention state of everything before it resident. A thread that has been alive for a week is a cache entry that never gets evicted. Multiply that by millions of free users and the memory bill stops being a rounding error.
That explains the specific shape of the limits. Expiring old threads frees cache. Capping input size bounds the cost of any single request. A daily quota caps the tail. Each of the three attacks the same line item from a different direction.

What it does not say
No official announcement, no published thresholds, no restoration date. That opacity is the part users are angriest about, and reasonably so: a metered free tier is a normal business decision, but one that arrives without notice reads as a malfunction until someone explains it.
There is also the paid question. Developers have reported finding quota-management code in the web front end, which has fuelled speculation about a consumer subscription. DeepSeek has never sold one, and its public position has been that the consumer product stays free while the API carries the revenue. Limits are the cheap version of that position; a paid tier is the expensive one.
What to do this week
Export anything you care about. Both the App and web client can dump a conversation, and a thread that cannot be resumed is a thread you can no longer search.
Move long-running work to the API. Token prices are unchanged and there is no daily cap: $0.15 per million input tokens on a cache miss at off-peak, $0.30 at peak, $0.60 to $1.20 per million output. A thousand conversations of moderate length cost less than a coffee, and nobody throttles you for using them.
Keep prompts short and contexts deliberate. The 40,000-character ceiling is generous for a question and tight for a pasted codebase. If you routinely paste a repository, you were already paying for it in latency.
The read
Compute scarcity is now a product constraint, not an infrastructure footnote. The company with the cheapest tokens in the market is rationing the free tier of its own chat product while its API stays open and unmetered by volume. That is a coherent position — the API has a price signal, the free App does not — but it does mean the free App is no longer a reliable place to keep a long project. Treat it as a scratchpad, and keep the archive somewhere you control.