LM Studio on Mac Studio: benchmarks and Continue setup

This tutorial tests a Mac Studio with a 14-core CPU, 40-core GPU and 128 GB of RAM as a local LLM server. The speaker installs LM Studio for macOS, selects developer mode and downloads GPT-OSS 120B in GGUF format. The loading walkthrough explains how to reduce context length if LM Studio warns that memory is insufficient.

For the demonstrated workload, the speaker reports roughly 150 watts of power use, about 70 GB of RAM usage and quiet fans. Reported generation speeds are 70 to 80 tokens per second for GPT-OSS 120B and 100 to 120 for the 20B model. These are observations from this setup; the video does not provide a measured Nvidia comparison. A separate example uses MedGemma 4B to describe a chest X-ray, illustrating image input with a compatible model.

The setup then turns LM Studio into a self-hosted server on port 1234. A MacBook Pro connects over the local network through Continue in VS Code. The configuration uses the model identifier from LM Studio and a matching context length. In AI agent mode, the demo asks questions about a Java application folder using @tree, checks dependencies and creates example.json after the user accepts the change. The speaker says the server answers without accessing the internet, demonstrating offline inference after the software and model downloads.