oMLX vs MTPLX vs LM Studio: Mac mini M6 LLM tests

The presenter compares oMLX, MTPLX and LM Studio on a machine identified as a Mac mini M6 with 32GB of RAM. The local LLM tests cover Gemma 4, Qwen 3.6 and Qwen 3.8 variants, plus Ternary Bonsai 2. Most models use four-bit quantization; Bonsai uses two-bit quantization. These are results from the presenter's setup, rather than general hardware guarantees.

Short chat prompts produce roughly 31 to 45 tokens per second with Gemma 4 26B A4B and reported readings of 58 to 65 with Qwen 3.6 35B A3B. Bonsai runs at about 11.5 to 11.8 tokens per second, with reported memory use around 16 to 20GB. Qwen 3.8 27B in MTPLX returns roughly 17 to 25 tokens per second.

The coding assistant task exposes limits that chat speed alone misses. Using Pi Coding Agent, the presenter requests a Playwright C# framework with NUnit, dependency injection and page objects. At a stated 65K context length, Qwen 3.6 hits a memory ceiling after about eight minutes. Gemma finishes in roughly three minutes but leaves empty classes and missing framework components.

Qwen 3.8 27B with Splash in LM Studio takes about 42 minutes. The presenter reports 26 to 28GB of memory use and demonstrates a successful .NET test run. He recommends 32GB for these workloads, while his coding tests show that this capacity still has context and memory limits.