Self-hosting DeepSeek
Self-hosting is not cheaper and it is not better. It solves specific problems. Start by working out which one you have.
Three questions first
1. Can your data leave your network?
No → you have to self-host, whatever it costs.
Yes → run the break-even numbers. Most teams find the API cheaper.
2. How much do you actually call it?
Under a few hundred thousand tokens a day → the API almost always wins.
Hundreds of millions a day → self-hosting may pay off, if utilisation stays high.
3. How big a model do you need?
Classification, extraction, translation → small models (7B–32B) run locally.
Complex reasoning, long context → you need the big model, and it will not fit on one machine.
Four ways to deploy
| Approach | Best for | Effort | Throughput |
|---|---|---|---|
| Ollama | Trying it out, small teams | Very low | Medium |
| llama.cpp | Single machine, CPU+GPU mix | Low | Medium |
| vLLM | Production, many concurrent users | High | High |
| Cloud GPU | Short peaks, no hardware | Low | High |
Two calculators
Read more
Choosing a quantization · Ollama setup · llama.cpp setup · Hardware guide · The real cost of self-hosting