Ollama

Ollama runs open models on your own hardware and exposes them over a small HTTP API. It is the engine behind the local lane: you serve a code-tuned open model with Ollama on a machine with a GPU, and opencode talks to it for the routine work (see local and frontier models ).

Serve a model

ollama serve                    # starts the API on port 11434
ollama pull <a-code-model>      # e.g. a 14B–32B coding model sized to your GPU
ollama run  <a-code-model>      # quick check that it answers

Pick a model that fits in your GPU’s memory — a 32B model needs roughly 20 GB, a 14B model about half that. A bigger model is slower but stronger; for the routine 80% the smaller end is usually enough, which is the whole point of putting this work on the local lane.

Serve once, use everywhere

Run Ollama on one GPU machine and reach it across your private network at http://<gpu-machine>:11434. The laptops and smaller boxes then use the local lane without each needing a GPU of their own — one card serves the whole set.

The agent that drives it: opencode .
The lane split: local and frontier models .

Source: content/tools/ollama.md · maintained in the nuilab-agenticai repository.

Serve a model

Next