What is DeepSeek?
DeepSeek is a Chinese AI lab whose open-weight and hosted models (DeepSeek-V3 chat, DeepSeek-R1 reasoning) offer OpenAI-compatible APIs at a fraction of the cost of GPT-4o.
Resources
Short guides, curated links, a glossary and FAQ for DeepSeek and its sister LLM providers.
Short guides
OpenAI-compatible endpoints
'OpenAI-compatible' means the provider speaks the same /v1/chat/completions wire format. Point the OpenAI SDK at their base_url, swap the model name, and you're done.
How prompt caching cuts your bill
DeepSeek (and several sister providers) cache the leading prefix of your prompt. Keep system prompts and few-shot examples stable to land near the cheap 'cache-hit' price rather than the 'cache-miss' price.
Reading a status page
A green dot means an endpoint is reachable. It does not confirm your key is valid, your model is available, or capacity is healthy — those live behind auth-gated internals we can't probe from a browser.
Why CORS limits browser checks
Provider APIs don't return permissive CORS headers, so browser checks confirm DNS/TCP/TLS handshake — not the actual JSON response. That's why we report reachability, not full health.
Curated links
Official status pages
Communities
Glossary
- Token
- Atomic unit of text billed by an LLM API — roughly 4 characters of English.
- Context window
- Max tokens (input + output) a single request can hold. DeepSeek hosted models: 64K.
- MoE (Mixture of Experts)
- Architecture where only a fraction of parameters activate per token, cutting inference cost.
- Distillation
- Training a smaller model on outputs of a larger one — DeepSeek ships R1-distilled Llama and Qwen variants.
- Chain-of-thought (CoT)
- Intermediate reasoning trace a model generates before the final answer. R1 exposes it as reasoning_content.
- OpenAI-compatible
- The provider's API speaks the same /v1/chat/completions wire format as OpenAI's SDK.
- Cache hit / miss
- Whether your prompt's leading tokens matched a recent prefix cache — determines input-token price on DeepSeek.
- TPM / RPM
- Tokens-per-minute and requests-per-minute rate-limit dimensions.
- SSE
- Server-Sent Events — the streaming transport used by chat completions when stream=true.
- Function / tool calling
- Structured JSON invocation of app-defined tools by the model.
- JSON mode
- Constrained output guaranteed to be valid JSON. Supported by most OpenAI-compatible providers.
- Fallback provider
- Secondary host of the same model used when the primary 429s or 5xxes.
- 429 Too Many Requests
- Rate limit hit — retry with exponential backoff.
- 402 Payment Required
- Prepaid balance exhausted — top up in the console.
- Router
- Client-side or gateway logic that picks the best provider per request and handles fallback.
FAQ
Is DeepSeek safe to use in production?
Yes for most non-regulated workloads. Review DeepSeek's data-retention policy and consider a self-hosted deployment (open weights on HF) if data residency matters.
Which OpenAI-compatible provider is cheapest?
For DeepSeek models specifically: DeepSeek direct. For Llama-class open weights: Together and Fireworks are competitive. Groq wins on latency, not raw price.
Why does this site show reachable but the official status page says incident?
Browser reachability checks only see whether public endpoints respond. Internal capacity, model availability, or account-scoped issues can degrade while the edge remains reachable.
Do you proxy or store my API keys?
No. deepseek.health never handles your API keys. All checks are unauthenticated and run from your browser.