When you run a local model, several things are involved: a client command, a long-running model service, model files, a runtime and CPU/GPU resources. Ollama hides much of that complexity, which makes it excellent for learning—but we should still understand the pieces.
Engineering focus: Local model service, process boundaries and Ollama mental model
When you run a local model, several things are involved: a client command, a long-running model service, model files, a runtime and CPU/GPU resources. Ollama hides much of that complexity, which makes it excellent for learning—but we should still understand the pieces.
Do not memorize product names first. Build the mental model, watch the system move, then map tools onto it.
The command you type is a client. The long-running service owns model execution and exposes the local API.
Ollama run or curl is a client request, not the model itself.
A CLI command is a client interface. The Ollama service is the long-running process that accepts API requests. The model is the set of weights and runtime behavior used to generate output. CPU or GPU hardware performs the computation.
Keeping these pieces separate makes errors easier to interpret. A missing model is different from a stopped service, and both are different from insufficient memory.
Closing a terminal client does not necessarily mean the model service stopped. Client lifecycle and server lifecycle are separate.
Ollama's local API is served by default at http://localhost:11434/api. `localhost` means the same machine as the client process. `11434` is the TCP port where the service listens, and `/api` is the API path prefix.
This becomes important when the client runs inside a container or on another machine: its `localhost` is no longer your Ollama host.
A connection-refused error is not an LLM reasoning failure. It is a service/network failure.
Once Ollama is running, a client can send JSON to its HTTP API. The request names a model and supplies input. The service returns generated data, often as a stream of JSON objects unless streaming is disabled for that endpoint.
Seeing the raw request is valuable because later gateways and applications are simply creating, forwarding and governing similar requests.
The important thing is to understand the boundary and evidence before memorizing implementation details.
When local inference fails, begin with the cheapest evidence: is the service process running, is the port reachable, does the API respond, is the requested model available, and only then inspect model/resource behavior.
Changing several things at once destroys evidence. Test one boundary at a time.
The important thing is to understand the boundary and evidence before memorizing implementation details.
curl http://localhost:11434/api/generate \
-d '{"model":"<your-model>","prompt":"Explain latency simply.","stream":false}'1. process → is the service running?
2. socket → is port 11434 reachable?
3. API → does HTTP respond?
4. model → is the requested model available?
5. resource → can it fit/run?Goal: Understand client/server/API boundaries with no dependencies, then map the same ideas onto a real local Ollama installation if available.
python3 public/labs/llmops/03_model_service_basics.py --self-testpython3 public/labs/llmops/03_model_service_basics.py --servecurl -s http://127.0.0.1:18080/api/generate -H 'Content-Type: application/json' -d '{"model":"learning-model","prompt":"hello"}'If Ollama is installed, inspect its local API with a model already present on your system. Compare the same client → service → model mental model.
The same client/server reasoning applies whether the backend is Ollama, vLLM, a managed provider or an internal inference service.
This checklist is a study aid, not a proficiency score. Your evidence record remains based on tracked learning activity and later formal assessment.
Answer before opening each explanation. Opening all four answers records knowledge-check activity in My Learning.
The Ollama API service. Clients such as curl or the CLI communicate with that service.
Each container has its own network namespace; localhost refers to the container itself, not automatically to the host.
Verify the service process, bind address, port and network path first.
The HTTP call exposes the actual client/server contract that applications, gateways and platform services will use later.
Progress is saved in this browser immediately. When you sign in to My Learning, it can also sync across devices.
Attempt the core lab and knowledge checks before marking the module complete.