Qwen3.6 with LM Studio and opencode: local coding setup

This tutorial uses Qwen3.6 27B, LM Studio and opencode to build a local AI coding assistant on a Mac. The speaker presents the model as capable of reasoning, image recognition and tool calling, and compares its agentic coding performance with Claude Opus 4.5. That comparison reflects the benchmarks and personal impressions presented in the video.

The setup covers enabling developer mode in LM Studio, downloading roughly 20 GB of model files and starting an OpenAI-compatible endpoint. The speaker then connects opencode through its JSON configuration, including image modalities. LM Studio's developer tab provides logs, context settings and concurrent sessions. The speaker describes cache quantization as a speed option with a possible quality cost.

Examples include identifying SF Symbols from a screenshot, generating Minesweeper and Snake, and adding a developer information card to an existing mobile app. The speaker reports that the larger task used 31 tool calls and took almost 10 minutes, with loud fans and substantial heat on a 96 GB M3 Ultra Mac Studio.

Hardware limits matter: the speaker recommends at least 24 GB of memory and reports failing to run the model on an 18 GB MacBook Pro. The coding examples are presented as working offline. A final phone demo connects Locally AI through LM Link to the computer; an attempted web search does not work.