oMLX vs Ollama and LM Studio: Mac coding setup

The presenter explains why they switched from Ollama and LM Studio to oMLX for a local LLM coding setup. On their 32GB Mac, they report that the other runners stopped responding or crashed after four or five prompts, which they attribute to context or RAM limits. This is an account of their setup, rather than a controlled comparison of the runners.

The walkthrough starts the oMLX server and opens its admin panel before moving to VS Code. The supplied description identifies the hardware as a MacBook Pro M5 and the coding assistant extension as Continue.dev. It names the model as Qwen3.6 35B-A3B in 4-bit form, roughly 19GB, though the transcript renders its name differently.

The demonstration asks the model to plan an app where developers can display their work and clients can hire them. With the model loaded, the presenter reports about 27GB of memory use and 1GB of swap. They describe SSD storage as helping with memory use; the clip does not isolate the effects of oMLX caching from macOS swap.

The model returns an app plan after a period of reasoning. The presenter reports generation at about 33 tokens per second and prompt processing at nine tokens per second. The example shows planning in VS Code, but does not demonstrate a completed app or test stability across a longer coding session.