Compare a local model with your baseline
Local execution changes the tradeoff
A local model can run on your own machine, but capacity, speed, and operational responsibilities depend on your hardware and the model. “Local” is a deployment choice, not a quality score. Check whether the full path stays local, including logs and optional cloud features.
Ollama provides tooling to run models and an API. Its official quickstart describes setup. Choose a model that fits your machine rather than treating a model name in a lesson as permanently current.
Check your understanding
What makes a fair local-model comparison?
Use the same cases and record actual measurements
The optional local-model runner in the lab bundle calls a local Ollama endpoint after you install and select a model. It sends one fictional dinner request and reports the actual elapsed time plus the response. It does not invent benchmark numbers or install software for you.
Run the same held-out requests with each candidate. Track field fidelity, invalid output, latency, and resource constraints. A fast response that changes the budget is not a successful optimization.
Check your understanding
Keep the live-model extension optional and concrete
Download the lab bundle, follow LIVE-MODEL.md, and use the provided command with your installed model name. If you are not ready to install a model, complete the decision memo using the clearly labeled scenario lab instead. Do not report simulated outcomes as measured model results.
Your final memo names the chosen deployment path, evidence, and remaining gaps. A beginner can make a sound product decision before becoming a model-training specialist.
Check your understanding
What would make you reject a faster model for this task?
Key takeaways
- Choose deployment based on evidence and constraints.
- Compare candidates on the same requests.
- Clearly label planned, simulated, and measured results.