Qwen3.8-27B UD-Q2_K_XL on Apple M5: a checksum-pinned llama.cpp benchmark
Five llama-bench repetitions measure prompt processing and token generation for one exact mixed-tensor GGUF, with the hardware, runtime, flags, and raw samples attached.
Reproducibility lab
Benchmarks with complete run metadata
Every result must disclose the model artifact, runtime commit, flags, hardware, prompt set, sample count, and raw artifacts.