Player not loading? Watch on YouTube
Leon demonstrates OpenCode with Ollama as a local coding assistant, then builds a Next.js chat app that uses Ollama for inference. He reports better results than with the same model in Claude Code. His fresh Claude Code session shows roughly 5,800 tokens for the system prompt and 20,000 for tools; the performance comparison reflects his experience rather than a general benchmark.
The setup covers installing OpenCode through npm, downloading Ollama, and using its pull and run commands. In OpenCode, he selects Ollama through the connect menu and enters Ollama as the API key. If a downloaded model is missing, he adds it to the OpenCode configuration.
His hardware recommendations include Gemma 4 E4B for about 8 or 12 GB of VRAM and Gemma 4 26B A4B for about 16 GB of RAM. For 24 GB of VRAM or more, the transcript names Code Llama 3.6, but later calls the demonstrated model Quinn 3.6. That naming inconsistency leaves the intended model unclear.
The workflow splits an implementation plan into six focused task files with acceptance criteria. Leon reviews the plan, removes a redundant scaffolding step, and starts a fresh conversation for each phase. The app initially fails to stream responses. He installs the agent-browser skill for testing, but reports that the agent uses Playwright instead. The final demonstration shows streamed replies and Markdown formatting. The tutorial gives a practical way to run models locally while showing why generated plans and working app behavior still need review.
