oMLX trails LM Studio on tokens/s

The likely answer is that this is a runtime-and-defaults problem, not a model problem. LM Studio’s MLX stack is mature and aggressively optimized for steady-state generation, while oMLX is optimized more for serving, persistence, and agent workflows than pure single-request tok/s.

• –oMLX’s own positioning emphasizes continuous batching, tiered SSD KV caching, and multi-model serving, which are useful for long sessions but can add overhead in a simple throughput test
• –LM Studio has a newer MLX engine path and a lot of polish around Apple Silicon performance, so small differences in backend settings can easily show up as double-digit token/sec gaps
• –“Same model” does not guarantee a fair comparison if context length, cache state, batch size, prompt format, or sampling defaults differ
• –For max Mac speed, the real benchmark is a clean apples-to-apples run with identical runtime, context, and decoding settings, not just the model file name
• –If the workload is coding agents or long-lived local serving, oMLX’s cache persistence may still be the better tradeoff even if raw decode speed is lower