Running an AI model: on your VPS or through an API?

What can a VPS without a graphics card do?

Without a graphics card the model runs on the processor and in memory.

• A compressed 7 to 8 billion parameter model needs about 5 to 6 GB of memory
• It answers slowly, a few words per second, and one request at a time keeps the server busy
• Large models (30 billion and more) do not fit in a common VPS

On a 12 GB VPS that also runs your site and database, little room is left.

When is the API the better choice?

As soon as you want quality, speed or several users at once.

• Access to much larger models (Qwen3 235B, DeepSeek, Claude, GPT-5)
• Fast streamed answers, several requests in parallel
• No memory or processor taken on your server
• You pay only for what you use

When to keep a model on your server?

For small, regular tasks, tests, or when data must never leave your server.

Then take enough memory (24 GB with XAB Business) and a small specialized model, and keep the API for heavy tasks.

The most common combination

The application runs on a managed VPS and the AI is called through an API. Your code stays the same: with an OpenAI-compatible API, moving from Ollama to XAB AI means changing the address and the key.

Every XAB Cloud Pro or Business server includes 5 TND of free XAB AI credit to test.