among the many things I found that were not highlighted anywhere (that I could find) is the distributed inference across Macs (JACCL over Thunderbolt 5), an OpenAI-compatible mlx_lm.server, agentic-on-Mac.

Apple keeps MLX (bring-your-own-weights) separate from Foundation Models / Core AI.