llama.cpp release build performance benchmarks on Apple M5 Max

RELEASE_BUILD_REPORT.md

llama.cpp Release Build Report

Date: 2026-04-12 Host: Apple M5 Max Repo commit: ff5ef8278

Final build

The final recommended build is in build-final.

Configuration:

• CMAKE_BUILD_TYPE=Release
• GGML_LTO=ON
• GGML_BLAS=ON
• GGML_BLAS_VENDOR=Apple
• GGML_METAL=ON
• GGML_METAL_NDEBUG=OFF
• GGML_NATIVE=ON
• GGML_OPENMP_ENABLED=OFF

Benchmark findings

Benchmark model:

• Ollama mistral:7b
• GGUF blob: /Users/jrepp/.ollama/models/blobs/sha256-f5074b1221da0f5a2910d33b642efa5b9eb58cfdddca1c79e16d7ad28aa2b31f

Benchmark mode:

• llama-bench
• device: MTL0
• GPU offload: -ngl 99

Prompt processing benchmark (-p 512 -n 0):

Build	Avg tokens/sec
stock build/	2126.74
build-lto/	2116.55
build-metalndebug/	1415.58
build-openmp/	1517.08
full tuned build-tuned/	1540.61
final build-final/	2118.91

Generation benchmark (-p 0 -n 128):

Build	Avg tokens/sec
stock build/	100.56
build-openmp/	99.57
build-metalndebug/	99.48
full tuned build-tuned/	86.32

Conclusions:

• GGML_LTO=ON is safe on this machine.
• GGML_METAL_NDEBUG=ON should be avoided for Metal-backed inference here.
• OpenMP should be avoided for this Metal-backed build here.
• The final build-final/ build preserves stock-level Metal performance while keeping LTO.

Validation

The final build completed successfully and these commands ran successfully:

• build-final/bin/llama-cli --help
• build-final/bin/llama-server --help
• build-final/bin/llama-bench -m <mistral_blob> --device MTL0 -ngl 99 -p 512 -n 0 -r 3 -o json

Notes

• Sandboxed runs could not access Metal correctly on this machine. Metal benchmarks only worked outside the sandbox.
• Existing Jinja-related compiler warnings are still present in the repo, but there were no build failures.