Gemma 4 12B-it Q4_K_M; Sherlock Holmes 8K-configured tier

Source-reported campaign measured 2026-06-05 by Kelly, published in SymbolStar/gemma4-mac-mini-bench; recovered 2026-10-13. This is newly recovered historical evidence, not a new run or local reproduction. Primary report: https://github.com/SymbolStar/gemma4-mac-mini-bench/blob/main/REPORT-longctx.md . Data: https://github.com/SymbolStar/gemma4-mac-mini-bench/blob/main/data/longctx-results.json . Short-task report/data: https://github.com/SymbolStar/gemma4-mac-mini-bench/blob/main/REPORT.md and https://github.com/SymbolStar/gemma4-mac-mini-bench/blob/main/data/results.json . Headline is one source row, the first long-context tier: 8.7 tokens/s decode and 110.4 tokens/s native prompt processing, 5713 actual input tokens and 512 actual output tokens, with num_ctx=8192. This is not an average, maximum, concurrency throughput or a full 8192-token prompt. The same-row source TTFT estimate is 64.5 s and request wall time 123.4 s. Hardware: one Apple Mac mini M4, Mac16,10, 10 CPU cores (4P+6E), 10 GPU cores, 32 GB unified memory; macOS 26.5.1. Ollama detected Metal and 25.0 GiB available in the unified pool; that available-memory value is not dedicated VRAM or installed RAM. Power, clock, temperature, storage model and GPU offload layer count are unknown. Model: unsloth/gemma-4-12b-it-GGUF, gemma-4-12b-it-Q4_K_M.gguf, GGUF Q4_K_M. The source says it downloaded Unsloth GGUF files and registered local aliases gemma4-e4b-q4 and gemma4-12b-q4 after the registry pull stalled. The generic README reproduction command uses the Ollama registry; do not substitute today's registry build for these measured Unsloth files. HF identity and currently available file were verified at https://huggingface.co/unsloth/gemma-4-12b-it-GGUF/blob/main/gemma-4-12b-it-Q4_K_M.gguf. Historical measured HF revision, file hash and complete Modelfile are unknown. Current HF main is not the tested revision. Runtime: Ollama 0.30.5 macOS cask build. Source flags: OLLAMA_FLASH_ATTENTION=1 and OLLAMA_KV_CACHE_TYPE=q8_0. Text-only local /api/generate calls, serial request loop, stream=false. Long-context request explicitly uses think=false, temperature=0.3, num_predict=512 and the listed num_ctx. Speculative decoding/MTP and other sampling settings are not reported and must remain unknown. Long-context workload: prefixes of Project Gutenberg's The Adventures of Sherlock Holmes plus a fixed Chinese structured-summary request. Input construction uses 3.5 characters per target token, while saved actual input counts come from Ollama prompt_eval_count; nominal targets are not measured token counts. Each of the listed tiers has one stored run. The script changes num_ctx per request. The report says KV allocation changes between tiers; no explicit cache flush, cached-token counter or independent cold/warm validation is provided, so do not label these as proven cold-prefill measurements. Native metric definitions verified in the published longctx.py: prompt_eval_count / (prompt_eval_duration in seconds) and eval_count / (eval_duration in seconds). It stores rounded PP to 1 decimal and TG to 2 decimals. The source's TTFT is (load_duration + prompt_eval_duration), a server-phase estimate from a nonstreaming request, not measured first-token arrival. PP is not input divided by TTFT or request wall time. Original unrounded phase durations are not retained in longctx-results.json. All long-context rows for this model (configured context; actual input/output; native PP/TG; source phase-TTFT estimate; request wall): 8192; 5713/512; 110.4/8.7 tokens/s; 64.5 s; 123.4 s 16384; 12707/512; 99.8/9.95 tokens/s; 129.35 s; 180.9 s 32768; 26279/489; 92.8/8.38 tokens/s; 284.78 s; 343.2 s Each row records swap_after=27 MB and swap_delta=0 MB. Thus there was no measured increase, not zero pre-existing swap. The script's mem_pct_after is a vm_stat non-free-page calculation and not a reliable macOS memory-pressure or model-allocation measurement; it is not used as VRAM/RAM capacity. Separate short-task campaign, same source setup and configured context8192, one warmup 'Say OK.' call per model then a2-second pause, then translation, SwiftUI coding and pandas analysis in sequence. The published bench.py omits explicit think/temperature options, so their measured settings remain unknown; do not copy long-context think=false into these rows. TG is eval_count/eval_duration. In each line below: task; actual input/output; prompt/eval durations in seconds; TG; load duration; source TTFT estimate; request wall. translate; 94/1100; 1.400089000/136.793086000; 8.04134209 tokens/s; 0.258447625 s; 1.658536625 s; 138.45989299 s swift_code; 74/2232; 0.871346000/267.384022000; 8.34754442 tokens/s; 0.265564333 s; 1.136910333 s; 268.54477787 s data_analysis; 174/1689; 1.975874000/215.092092000; 7.85245048 tokens/s; 0.259922250 s; 2.235796250 s; 217.35145402 s Short-task prompt rates can be recomputed from the retained native counts/durations, but are not combined with the long-context headline. Short-task 'peak_rss_gb' is a post-request ps RSS sample over matching processes, not a continuously monitored GPU peak. Repository provenance: data files show commit f6cffa8f4680ce1e0dcab8b21ee65b281c88d2ce (June6). Published harness currently shows 2bc218b3fa7b931916f837a5b3cb406571d1155a, whose commit description says it sanitizes local paths. The historical executable/harness revision is not pinned to the run, and the saved preview length differs from the current script; use the published script to establish the reported metric method, not a claim of byte-for-byte reproduction. Model quality is not independently validated by this throughput campaign.