Mac LLM Bench
A community-contributed benchmark database of language-model speed and code-quality results measured on Apple Silicon Macs, organized by chip generation and configuration so someone can look up expected performance before choosing a model.
Mac LLM Bench compares two runtimes under fixed, documented conditions so that submitted results are comparable to each other: llama-bench for GGUF models quantized with Q4_K_M through llama.cpp, and mlx_lm.benchmark for MLX models with 4-bit quantization, both measuring prompt-processing and text-generation speed at fixed token counts with all GPU layers offloaded and flash attention enabled. A separate code-quality benchmark uses EvalPlus's HumanEval+ suite against either runtime, and results for GGUF and MLX are displayed side by side with a runtime column so the two can be compared directly.
Results are organized by Apple chip generation, from M1 through M5, and within each generation by variant, such as base, Pro, Max or Ultra, and hardware configuration such as CPU core count, GPU core count and RAM; several generations are still awaiting community-submitted results. The project notes that some newer models may not yet be supported by every runtime, since MLX support in particular depends on the version of the mlx-lm library installed.
It is written in Shell scripts and released under the MIT license.