How to run a chat request with the Ollama API

The api/chat endpoint accepts role-based messages and returns an assistant message, which makes it the normal API path for chatbots and multi-turn application flows. It preserves conversation structure instead of flattening everything into one prompt.

Use messages for user, assistant, and tool entries. Set stream to false for a compact smoke test, then enable streaming after the caller can parse the final response reliably.

Thinking-capable models may include a separate thinking field. Application code should display or store that trace deliberately rather than treating it as the final answer.

1. Send a minimal chat request with one user message.

   $ curl -s http://localhost:11434/api/chat -d '{
     "model": "gpt-oss:20b",
     "messages": [{"role":"user","content":"Reply with OK."}],
     "think": "low",
     "stream": false,
     "options": {"num_predict": 24}
   }'
   {"message":{"role":"assistant","content":"OK","thinking":"User wants \"OK\"."},"done":true}

2. Read the assistant answer from message.content.

   $ python3 - <<'PY'
   payload = {"message": {"role": "assistant", "content": "OK"}}
   print(payload["message"]["content"])
   PY
   OK

3. Send prior assistant messages when the next turn depends on context.

   $ curl -s http://localhost:11434/api/chat -d '{
     "model": "gpt-oss:20b",
     "messages": [
       {"role":"user","content":"Remember the code word basalt."},
       {"role":"assistant","content":"OK"},
       {"role":"user","content":"What code word did I give?"}
     ],
     "stream": false
   }'
   {"message":{"role":"assistant","content":"basalt"},"done":true}

4. Use keep_alive when the next request should reuse the loaded model.

   $ curl -s http://localhost:11434/api/chat -d '{"model":"gpt-oss:20b","messages":[{"role":"user","content":"OK?"}],"stream":false,"keep_alive":"10m"}'
   {"done":true}

   Related: How to set Ollama model keep alive

5. Check loaded processes after the chat request.

   $ ollama ps
   NAME           ID              SIZE     PROCESSOR    CONTEXT    UNTIL
   gpt-oss:20b    17052f91a42e    12 GB    100% GPU     131072     10 minutes from now

   Related: How to list running Ollama models