Ollama model routing on laptops

The real takeaway is that one local model will not fit every agent step, so orchestration matters as much as model choice. Ollama can serve different models via its API, but the intelligence for picking the right model per task usually has to live in the agent layer.

• –The post captures the core local-LLM tradeoff: 20B-class models run out of room, 7B models can be too weak, and mid-size models often hit the best compromise.
• –For agentic workflows, trivial probes, heartbeats, and classification steps should usually be routed to a smaller, cheaper model or even a rules layer.
• –Context exhaustion is a separate issue from model size; trimming history, shortening prompts, and keeping task-specific state outside the chat window matter just as much.
• –This is a strong fit for Ollama because its API is built around selecting a model per request, which makes model routing feasible even if Ollama itself does not decide the policy.