Pat Simmons starts with local inference, using Ollama to download and run a compact vision model on his own computer. He then launches Qwen 3 Coder inside an agent harness and shows it searching a private Obsidian vault through an explicit skill, illustrating how a smaller local model can handle narrow, well-defined workflows without sending the underlying notes to a hosted model.
For larger models, Simmons demonstrates NVIDIA NIM through OpenCode. The hosted service provides access to several open models without local hardware requirements, although he notes that availability, speed and usage limits depend on NVIDIA's service. A coding demo completes after he switches away from a model that remains unresponsive.
The final setup uses OpenRouter as a single API route for multiple open and closed models. Simmons configures an agent harness to call GLM 5.2 and compares its generated site with an earlier MiniMax result. The demonstrations are practical walkthroughs rather than controlled benchmarks, and model names, prices, access terms and availability can change. The creator's bootcamp and newsletter promotions are omitted.
Watch on YouTube
