Just as this domestic model was really starting to improve, it ran into problems.
1 Like
Both the client and server are busy, so it’s unavailable.
The WB reverse proxy still works.
I totally agree—DeepSeek’s API goes haywire every other day, so I stopped relying on it being stable a long time ago. In practice, there are three things you can do:
First, set up a basic health check that periodically curls /v1/models. Trigger an alert on a timeout or a 5xx response; at least you’ll know right away whether it’s their service that’s down rather than a problem with your code.
Second, make BASE_URL configurable in your code and keep two sets of keys for primary and backup services. If the official service goes down, switch directly to a relay compatible with the OpenAI API format. The interface is the same anyway, so you just need to change an environment variable and restart.
Third, add a circuit breaker and exponential backoff retries to your requests. Don’t let them hang indefinitely and bring down the entire chain.
For lightweight tasks, you can also run a small model locally with Ollama as a fallback. Set things up this way, and you’ll barely notice when DeepSeek goes down again.