LM Studio tops oMLX in M1 Ultra test

The hot take: this is less a shocking upset than a reminder that Apple Silicon benchmarks are brutally workload-specific. oMLX is optimized around agent-style cache reuse and batching, while LM Studio can shine on straightforward single-stream GGUF runs. LM Studio supports both llama.cpp and MLX backends, and its MLX support was added specifically for Apple Silicon, but GGUF on llama.cpp can still be very competitive for raw single-user generation speed. oMLX positions itself around paged SSD KV caching, continuous batching, and long-session agent workloads, so its headline advantage is often lower repeated-prefill pain and better concurrent throughput, not always best one-shot tok/s. A 2025 Apple Silicon runtime study on an M2 Ultra found MLX had the highest sustained throughput under its test setup, while llama.cpp stayed highly efficient for lightweight single-stream use; both can be right depending on benchmark design. If the Reddit test mixed model formats or different quant builds, it is not apples-to-apples. MLX and GGUF conversions of the same model can differ in tokenizer handling, kernel maturity, memory pressure, and effective context settings. For developers, the practical takeaway is simple: benchmark TTFT, prompt ingest, steady-state tok/s, and repeat-turn performance separately. The fastest stack for chat UI use is not automatically the fastest stack for coding agents or long-context sessions.