dsh-mlx-local

Running a local LLM on an Apple Silicon Mac used to mean hand-rolling a Python venv, babysitting mlx_lm.server in a terminal, and then figuring out how to wire the endpoint into DeepSeek Harness as a provider. dsh-mlx-local collapses all of that into a single plugin: it auto-detects a compatible Python (3.9-3.13), spins up an isolated venv, installs mlx-lm, and exposes a dedicated MLX panel in Settings where you start, stop, and switch models with one click. It also watches for abnormal exits, reaps orphaned processes, and shuts the server down cleanly when DSH quits, so a zombie is never left holding port 8080.

Model management is built in: add or remove local model directories, pre-download Hugging Face weights, and list what is already on disk. Switching models gracefully stops the old service before launching the new one. For thinking-capable models like Qwen3, the main model picker automatically surfaces a thinking-strength toggle (Off / High), removing the need for per-model manual configuration. The plugin does not register any extra provider; the local endpoint plugs straight into DSH's existing custom-provider mechanism via its OpenAI-compatible route at 127.0.0.1:8080/v1.

If you have an M-series Mac, want inference to stay on-device instead of round-tripping to the cloud, and would rather have a working local model in ten minutes than spend an evening reading docs, this plugin is built for you. As long as you are on macOS 13 or later and have a few GB of headroom for a 4-bit checkpoint, you are ready to go.