Mac Studio as Homelab AI Node

I bought a used Mac Studio M1 Max for all my LLM inference. Running qwen3.5:35b-a3b-coding-nvfp4 and gemma4:26b through Ollama now.

The wiki pipeline runs three models in sequence: gemma4:e2b for cleaning, qwen3.5:35b for crystallization, minicpm-v:8b for PDF pages. Although I got a few more in rotation. All three stay in memory between calls. The n8n pipeline calls Ollama for all synthesis. Every hour, metrics from Prometheus, Uptime Kuma, UniFi, and the Synology get handed to Ollama in parallel for summarization. The nomic-embed-text model also runs on this machine, with embeddings stored in pgvector for semantic search across the wiki, but thats half a gig.

64GB unified memory is nice. The wiki pipeline runs in a fraction of the previous time. I might even go teeny tiny and run concurrently.

Also I’ve noticed. Since I’m using ollama lol, the mlx models are more limited, but I’m running qwen3.5:35b instead of 3.6. It doesn’t freeze as often but is dumber. gemma4:26b is pretty solid. 10-second settle after vision calls with minicpm-v. but only because it gives it tiime to rest.

The Pavilion now handles Jellyfin with MX550 NVENC. Cleaner split.