Qwen3.8-27B UD-Q2_K_XL on Apple M5: a checksum-pinned llama.cpp benchmark

What this run answers

On the recorded Apple M5 system, the checksum-pinned Qwen3.8-27B GGUF averaged 39.439795 tokens per second for prompt processing and 1.865259 tokens per second for token generation. Both figures are means from five llama-bench repetitions. The prompt workload used 512 tokens; the generation workload used 128.

The system was a MacBook Air with Apple M5, 10 logical cores, 24 GB memory, and macOS 26.6.2 (25G83). The run used llama.cpp llama-bench build 10470 at commit 34af94cd9, with BLAS and the Apple M5 Metal backend. Arguments were --n-prompt 512, --n-gen 128, --repetitions 5, --threads 10, --n-gpu-layers 99, --batch-size 2048, --ubatch-size 512, and --flash-attn auto.

Those numbers describe one file, machine, runtime build, and argument set. They are not a general ranking for Qwen3.8-27B, Apple Silicon, or every file carrying the same distribution label.

The artifact is pinned, not inferred from its filename

The embedded GGUF header identifies Qwen3.8-27B, names Qwen as the base-model organization, and links to the Qwen/Qwen3.8-27B repository. The local file, with SHA-256 46151b52a5cad673d90a00222103254864326c251130b8fc4381d6f34386b3c8, also matches the immutable Unsloth revision f1bfb127c64f7072bdd2cad55f258b9c8b2910fe by filename, byte size, and LFS SHA-256. The pinned model card declares Apache-2.0. This is publisher metadata, not legal advice.

Why the quantization labels look different

Unsloth calls this mixed-tensor distribution UD-Q2_K_XL. Its GGUF metadata stores general.file_type = 14, which llama.cpp renders as Q4_K - Small. A verbose load of the exact artifact counted F32, Q2_K, Q3_K, IQ3_XXS, IQ3_S, IQ4_XS, and IQ1_M tensors. The two labels describe the artifact at different levels; the llama.cpp aggregate label does not mean that every tensor is Q4_K_S.

Limits and reproduction boundary

The run record does not contain power mode, thermal state, room conditions, or energy use. It uses synthetic token workloads, so it does not measure answer quality or application-level task performance. The recorded arguments do not include --n-ctx or another explicit context-length flag, so this page does not state a context size.

The raw result, individual samples, model checksum, llama.cpp commit, hardware record, and complete argument list are linked on this page. A reproduction should preserve those inputs before treating any difference as a performance change.