Gemma 4 12B

Source: https://github.com/HDZucht/Mac-Mini-M6-LLM-benchmarks Raw measured rows: https://github.com/HDZucht/Mac-Mini-M6-LLM-benchmarks/blob/3990a8e0fa5f2762911689439ce12e1c28d4c989/data/speed_2026-10-03.jsonl Hardware/method/model labels: https://github.com/HDZucht/Mac-Mini-M6-LLM-benchmarks/blob/cb99870e05aa9f5783c5c6987ddd9154834123ab/README.md Loaded context metadata: https://github.com/HDZucht/Mac-Mini-M6-LLM-benchmarks/blob/main/data/loaded_context_2026-10-03.jsonl Harness: https://github.com/HDZucht/Mac-Mini-M6-LLM-benchmarks/blob/main/bench_speed.py Original author: HDZucht / Hans-Dieter Zucht, with repository-credited automated measurements by Argus. Campaign dated October 3–4, 2026; these exact records are in the October 3 raw file. No local reproduction was performed. Headline: 18.8 tokens/s is the median of three English-prose runs (Ruhr economy). It is client-observed streaming decode throughput, not a pooled rate, overall mean, peak, native kernel result or prompt-processing rate. Five workload medians and all 15 runs are preserved separately below. Hardware: Apple Mac mini M6 (Mac18,5), 12 CPU cores (2 Super, 4 Performance, 6 Efficiency), 12 GPU cores, 32 GB unified memory, macOS 27.0.1. One machine, one model loaded alone. Model storage: ORICO 2 TB external NVMe in an ASMedia ASM246x enclosure, source-reported 40 Gb/s PCIe 4.0 x4. Apple automatic fan control. Per-run temperature, clock, power, swap, exact GPU offload and peak resident memory were not logged. Exact measured API alias: google/gemma-4-12b-qat. Display model: Google Gemma 4 12B QAT. The author names the dense Gemma 4 12B QAT model; its QAT label is preserved separately from its unknown quantization. Source Format-column label: QAT. Exact measured quantization: unknown. Measured weight format: unknown; source label is not a confirmed weight container. Application: LM Studio. Backend: unreported for this alias. Exact app/backend versions are unknown. Do not turn QAT, MLX or Splash labels into an assumed Q4 quantization. The current exact LM Studio alias links the verified QAT GGUF collection. Current availability is not historical run metadata. The benchmark gives QAT only, so measured format/backend and quant code remain blank. Current alias corroboration (not original run metadata): https://lmstudio.ai/models/google/gemma-4-12b-qat Verified public related package/collection: https://huggingface.co/lmstudio-community/gemma-4-12B-it-QAT-GGUF Measured Hugging Face repository, filename, revision and hash are unrecorded. Public related-package links above prove availability only. Loaded context for this exact alias is 32,768 tokens. Requested load context was 32,768. Loaded capacity is not actual prompt length or a filled-context test. Actual input token counts are absent. Native prefill throughput/duration are unknown; keep PP and new-input-token fields blank. Method: one stream, max_tokens 600, temperature 0.7, reasoning_effort="none", three runs per prompt. A short Hallo warm-up precedes the measured prompts. The script includes emitted content and reasoning chunks; requested reasoning-off does not establish that reasoning was disabled in every backend/model. The harness computes (tokens - 1) / (stream_end - first_nonempty_content_time). It uses API completion_tokens when available and nonzero; otherwise it falls back to the number of nonempty streamed chunks. Raw records contain tokens, chunks, TTFT and generated text, but not the original usage object, finish_reason or total duration. Treat tokens as recorded output counts, with this usage/chunk fallback caveat. Chunk counts need not equal tokenizer token counts. Rates are the saved rounded client rates, not recalculated from TTFT. TTFT is end-to-end client latency, not native prefill duration. Prompts repeat identically across three repetitions; no nonce or explicit cache clear appears in the harness. Prefix-cache state/cached-token counts are unknown. These are not verified cold-prefill runs. A record of 600 equals the requested cap numerically, but finish_reason is not available. KV dtype, batching, microbatching, thread counts and exact GPU allocation are unreported. All 15 runs follow, in source repetition order. Fields: decode tokens/s; recorded output count; nonempty chunks; client TTFT seconds. German prose, Rhine geology (wissen_de), median 18.2 tokens/s: Run 0: 18.2 tokens/s; 459 output count; 458 chunks; TTFT 0.43 s. Run 1: 18.6 tokens/s; 423 output count; 422 chunks; TTFT 0.14 s. Run 2: 18.2 tokens/s; 426 output count; 425 chunks; TTFT 0.14 s. German prose, bead-based immunoassay (beads_de), median 18.5 tokens/s: Run 0: 18.5 tokens/s; 563 output count; 562 chunks; TTFT 0.25 s. Run 1: 17.7 tokens/s; 520 output count; 519 chunks; TTFT 0.17 s. Run 2: 18.7 tokens/s; 514 output count; 513 chunks; TTFT 0.16 s. English prose, Ruhr economy (prosa_en), median 18.8 tokens/s: Run 0: 18.7 tokens/s; 345 output count; 344 chunks; TTFT 0.25 s. Run 1: 18.8 tokens/s; 352 output count; 351 chunks; TTFT 0.14 s. Run 2: 18.9 tokens/s; 350 output count; 349 chunks; TTFT 0.14 s. Python ISO-8601 duration parser (code_py), median 18 tokens/s: Run 0: 18 tokens/s; 600 output count; 600 chunks; TTFT 0.31 s. Run 1: 18.7 tokens/s; 600 output count; 600 chunks; TTFT 0.15 s. Run 2: 17.1 tokens/s; 600 output count; 598 chunks; TTFT 0.17 s. JSON antigen/disease extraction (json_extr), median 18.2 tokens/s: Run 0: 18.2 tokens/s; 85 output count; 84 chunks; TTFT 0.36 s. Run 1: 18.2 tokens/s; 85 output count; 84 chunks; TTFT 0.13 s. Run 2: 18.4 tokens/s; 85 output count; 84 chunks; TTFT 0.13 s. Quality scope: The README reports the JSON task passed 3 of 3 automatic checks for this model. Prose/code quality was not generally graded. Throughput is not a quality score.