Ollama setup: Gemma 4 vs Qwen 3.6 with OpenCode

This tutorial tests whether a local LLM can handle everyday work through Ollama. The speaker uses an M4 Max Mac Studio with 16 CPU cores, 40 GPU cores, 64 GB of memory and 1 TB of storage. He accesses it remotely from a MacBook Pro through Tailscale and SSH, so the demonstrated workflow depends on a network connection.

The setup covers downloading models, listing installed models and stopping loaded models to release resources. It also explains quantization and the distinction between total and active parameters in mixture-of-experts models. In the demonstrated installation, Gemma 4's latest tag points to E4B.

On a shared question about Labor Day, the speaker reports about 78 tokens per second for Gemma 4 E4B, 45 for Qwen 3.6 35B-A3B and 16 for Qwen 3.6 27B. He prefers the 35B-A3B response as a compromise between speed and detail. These results describe his hardware and prompts, rather than general performance guarantees.

The OpenCode coding assistant demonstration uses Qwen 3.6 to explain linked lists and write a JavaScript implementation. The speaker cites a recommended context window of at least 64,000 tokens. A web-fetch task needs explicit instructions and URL corrections. He considers simpler programmatic tasks plausible uses for local AI, but still favors stronger cloud models for professional programming.