oMLX vs LM Studio: Apple Silicon setup and coding test

This walkthrough compares oMLX with LM Studio for a local LLM coding task on an M2 MacBook Pro. The presenter describes oMLX as a runtime built on Apple's MLX framework, with a two-tier KV cache that keeps immediate context in unified memory and stores older cached context on SSD.

The setup covers launching the server, entering an API key and downloading a 4-bit Qwen 3.6 model with 35 billion parameters. The presenter connects Codex CLI as a coding assistant and asks it to build a movie search app with a watchlist and ratings. He chooses Codex over Claude Code because he reports lower initial context overhead in this setup.

In the reported test, oMLX takes roughly 20 minutes and averages about 47 tokens per second, with 89% cache efficiency. It also encounters context-limit 400 errors that require session clearing. The presenter attributes faster processing of repeated prefixes to persistent SSD caching. The resulting app supports movie searches and ratings, but its watchlist resets after a page refresh.

LM Studio takes about 35 minutes at 16 tokens per second in the comparison. The presenter reports heavier memory pressure and difficulty multitasking, though it produces no context-limit 400 errors. These results describe one task and configuration; they do not establish a universal performance advantage.