LM Studio, VS Code and Continue: Mac coding setup

This tutorial builds a local AI coding workflow on an M-series Mac using LM Studio, VS Code's built-in chat and Continue. The speaker explains model size, quantization and unified memory before choosing Qwen 3.6. Model selection includes checking tool use, reasoning support and available memory, with room left for macOS, other apps and context.

The setup uses MLX files in LM Studio. The speaker recommends this format for Mac performance, but reports 39 tokens per second in one test against a tool's estimate of 137. Those figures describe this run, not a general speed guarantee. For the coding demonstration, the speaker loads a smaller 9-billion-parameter Qwen model with 128,000 tokens of context, then connects VS Code through a custom chat completions endpoint. The configuration assigns 100,000 tokens to input and 28,000 to output.

A separate Qwen Coder 2.5 MLX model supplies inline suggestions through Continue. The speaker's machine has 32 GB of memory; the speaker describes the selected autocomplete model as under 1 GB, without measuring its full runtime footprint. The AI agent creates a single-file "Would You Rather" game, and Continue repeats the task after a planning step. The closing section connects Hermes Agent to the same server and configures Kilo Code with LM Studio as a custom provider. Kilo Code's optional sign-in models are presented separately from that local path.

The speaker presents the workflow as useful for boilerplate and small web apps, while warning that complex architecture and deep debugging remain harder. Larger projects may need smaller tasks, and an 8 GB Mac limits model choice.