If the DeepSeek API is slow at 3 a.m. Beijing time, the likeliest reason is simple: that hour is cheap, so it is crowded. DeepSeek bills night hours in China at half price, and those hours are the working day in Europe and the Americas, so “the middle of the night” is rush hour for its servers.
The question came from a developer on the Chinese forum V2EX on 22 October 2026 (Beijing time). At about 3:40 a.m. they said DeepSeek’s official API had been stuck at roughly 60 tokens per second for two hours.
Key takeaways
- One user measured about 60 tokens per second (tps) at 3:40 a.m. Beijing time. Another user in the same thread reported about 257 tps.
- DeepSeek’s status page logged no incidents between 15 and 22 October 2026, so this looks like load, not an outage.
- Since August 2026, DeepSeek bills off-peak hours at half the peak rate. Beijing’s night falls inside that cheaper window.
- A new model, DeepSeek-V4.1-Flash-1017, went live on 17 October 2026, four days before the complaint.
- Per-user speed drops when servers pack more requests together. Your speed also depends on how you measure it.
What did the developer actually see?
The original post is short. It names no model, gives no request sizes and shows no error codes. It only says output had crawled at about 60 tps for two hours, ending around 3:40 a.m. on 22 October 2026 Beijing time, which was 19:40 UTC on 21 October.
Tokens per second is how fast the model writes once it starts. A token is a chunk of text, roughly three-quarters of an English word. At 60 tps, a 2,000-token code file takes more than half a minute to stream. At 257 tps, it takes about eight seconds.
The reply reporting about 257 tps came from someone using DeepSeek Harness, the company’s own coding-agent tool. Nobody in the thread offered an explanation. That gap between two users at the same hour is itself a clue.
Why is the DeepSeek API slow at night in China?
DeepSeek has spent more than a year steering traffic into China’s night. In February 2025 it cut prices by up to 75% between 00:30 and 08:30 Beijing time, a window that TechPartner News noted covers daytime in Europe and the US.
In August 2026 the scheme became time-of-day billing. According to API gateway APIYI’s summary of DeepSeek’s notice, from 16:00 UTC on 16 August 2026 off-peak tokens cost half the peak rate. DeepSeek said the change was meant to “allocate resources more reasonably”.
| Window | Beijing (UTC+8) | UTC | New York (EDT) | Billing |
|---|---|---|---|---|
| Peak 1 (weekdays) | 09:00-12:00 | 01:00-04:00 | 21:00-00:00 | Full price |
| Peak 2 (weekdays) | 14:00-18:00 | 06:00-10:00 | 02:00-06:00 | Full price |
| The V2EX slowdown | about 01:40-03:40 | about 17:40-19:40 | about 13:40-15:40 | Half price |
| Everything else, plus weekends | remaining 17 hours | remaining 17 hours | remaining 17 hours | Half price |
Look at the slowdown row. It landed in the cheap window, during the American working afternoon. Think of it like a toll road that is free after midnight: the lorries all wait for midnight, and the “empty” road jams.
Cheap hours also suit batch work: overnight evaluations, data labelling and long agent runs that nobody watches. Those jobs are exactly the kind teams schedule for the half-price window.
Did the new V4.1-Flash model add to the load?
Possibly. DeepSeek released DeepSeek-V4.1-Flash-1017 on 17 October 2026, served under the deepseek-flash name and tuned for coding agents. It also added Search and Fetch APIs that are free through DeepSeek Harness, Claude Code and Codex until billing starts on 31 October 2026.
Free tool calls and a fresh model tend to pull in agent traffic. Agents are greedy: one coding session can fire hundreds of long requests. DeepSeek has not published any load figures, so treat this as a reasonable guess, not a confirmed cause.
Why can two users get 60 and 257 tokens per second?
Language model servers batch requests: one GPU pass produces the next token for many users at once. Bigger batches mean more total output and cheaper tokens, but each user gets a smaller slice. When a provider is busy, it can keep everyone connected by growing the batch, and individual speed falls.
Beyond load, several things change the number you see:
- Prompt length. A 200,000-token context makes every new token more expensive to produce than a short chat does.
- Measurement. Some tools count time to first token in the total. Others only time the streaming part. The two can differ a lot.
- Hidden reasoning. If thinking tokens are not shown, visible output looks slower than the model really is.
- Network. Distance and routing to DeepSeek’s servers add delay, especially on long streams.
My take: a single forum post is not proof of throttling. But a slow afternoon in New York hitting a half-price window is exactly the outcome the pricing is designed to create.
What can you do when DeepSeek slows down?
- Check the status page first. DeepSeek’s status page showed 99.47% uptime for the V4.1 Flash API from July to October 2026. If there is no incident, you are probably seeing load.
- Measure properly. Log time to first token and output tokens divided by streaming time, separately.
- Know your limits. The rate limit docs set 2,500 concurrent requests per account for
deepseek-flashand 500 fordeepseek-v4-pro. Over that you get HTTP 429 errors, not slower output. Expansion requests are free. - Keep a fallback. V4.1-Flash weights are open under the MIT licence, so third-party hosts can serve it. A gateway with routing helps; see our piece on Vercel AI Gateway session affinity.
- Move batch jobs. If the job is not urgent, try a different off-peak slot and compare.
What happens next
Two dates matter. Search and Fetch billing begins at 16:00 UTC on 31 October 2026, which may cool some free agent traffic. DeepSeek has also said V4-Pro traffic runs on V4.1-Flash until a V4.1-Pro arrives, with no date given, according to The Next Web.
The bigger trend is that time-of-day pricing will keep shaping speed. If you rely on DeepSeek for production, it is worth watching for policy news too, like the DHS move to pull DeepSeek from US agencies.
FAQ
Is the DeepSeek API down when it gets this slow?
Not necessarily. DeepSeek’s status page showed no incidents from 15 to 22 October 2026. Slow streaming with no errors usually points to heavy load, not an outage.
When are DeepSeek’s peak hours?
Under the August 2026 scheme, peak is 01:00-04:00 and 06:00-10:00 UTC on weekdays. All other hours, and weekends, are off-peak at half price.
Does paying peak prices get me faster responses?
DeepSeek does not promise that. The price depends on the clock, not on a speed tier. Speed depends on how busy the servers are at the time.
Next time your tokens crawl, log the timestamp in UTC and your measured tps. A week of data will tell you more than any forum thread.
1 comment