Is DeepSeek down?
DeepSeek's live incidents are below. During an outage, the API returns HTTP 503 "Server Overloaded" or HTTP 500, or holds the connection without answering. Retry after a short wait, switch model, or run the same DeepSeek model on another host: the weights are open.
Live DeepSeek status
Checked
No open incidents
- Last incident: DeepSeek Web/API Degraded Performanceresolved Oct 12
Source: status.deepseek.com, DeepSeek's incident feed. Read on our server about every 2 minutes.
Which error is DeepSeek, and which is you
| What you see | What it means |
|---|---|
| HTTP 503, "Server Overloaded" | High traffic at DeepSeek. Retry after a brief wait |
| HTTP 500, "Server Error" | An issue on DeepSeek's servers. Retry, then check the status page |
Empty lines or : keep-alive comments, no answer | DeepSeek holds the connection while it waits to start, and closes it if inference has not started after 10 minutes |
| HTTP 402, "Insufficient Balance" | Your DeepSeek balance ran out |
| HTTP 429 | You are sending requests too fast for your account |
Five minutes into a DeepSeek outage
- Check whether the incident names the API or only the chat service: DeepSeek reports them as separate components.
- Retry with backoff, and set a client timeout so a held connection does not stall your agent for minutes.
- Switch to another model with the steps below, or run the same DeepSeek model on another host that serves the open weights.
How to switch model in your agent
| Agent | Steps |
|---|---|
| Claude Code on DeepSeek | Point ANTHROPIC_BASE_URL at another provider and restart, or unset it to go back to Anthropic |
| Cline, Roo Code, Kilo Code | Open the extension's settings, change API Provider or Model, and save. The next request uses it |
| OpenCode | Type /models and pick a model from another provider, or start with opencode -m provider/model |
| Aider | Type /model followed by another model name; the chat history stays |
| Your own code | Catch 5xx and 529, retry with exponential backoff, then send the request to a fallback model |
What Standard Compute does when DeepSeek fails
- Every request goes through OpenRouter. Standard Compute serves its DeepSeek models from other hosts of the same open weights, in a reviewed order per model, rather than from DeepSeek's own API, so an outage limited to api.deepseek.com does not stop those hosts.
- If the preferred host fails, OpenRouter can send the request to another host of the same DeepSeek model.
- Each failed call is retried twice on a 429, a 5xx or a network error, after 0.5 and then 1 second, or after the provider's Retry-After up to 8 seconds. If every attempt fails, your agent gets HTTP 503 with "The model is temporarily unavailable. Please retry shortly."
- A specific model ID always stays that model; failover only changes the provider. Only the auto-router, the model
standardcomputewith no model selection, can switch to a different model.
The catch
Other hosts run the same open weights, several at FP8 precision, and a new DeepSeek release can reach them later than DeepSeek's own API. Test a host on your own tasks before relying on it.
Where Standard Compute fits
Standard Compute is an OpenAI- and Anthropic-compatible LLM gateway for coding agents at https://api.stdcmpt.com/v1, with monthly plans from $19 (USD) and a fixed compute budget. One key serves Claude, GPT, Gemini, DeepSeek and GLM models, so moving off a struggling model during an outage takes one changed model ID. DeepSeek V4.1 Flash is deepseek/deepseek-v4.1-flash and V4 Pro is deepseek/deepseek-v4-pro.
FAQ
Is DeepSeek down or is it just me?
Check status.deepseek.com, which lists the API and the chat service separately. If no incident is open, read the status code: 401 is your key, 402 your balance, 429 your request rate, and 503 DeepSeek's capacity.
Why does the DeepSeek API hang without an error?
Under load DeepSeek keeps the request open and sends empty lines, or keep-alive comments when streaming, until inference starts. If it has not started after 10 minutes, DeepSeek closes the connection. Set a shorter client timeout and retry.
Can I use DeepSeek models when DeepSeek's own API is down?
Yes. DeepSeek publishes open weights, and other providers host them. Standard Compute serves DeepSeek V4.1 Flash and V4 Pro from such hosts rather than DeepSeek's own API, with failover between hosts of the same model.
Is DeepSeek on another host the same model?
The weights are the same, but precision can differ: several hosts run FP8. Testing a host on your own tasks is the reliable check.