Qwen3.8 27B Splash

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: 20.2 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: qwen3.8-27b-splash. Display model: Qwen3.8-27B. The author separates this Splash alias from Qwen3.8 GGUF and Swift-Qwen3.8 Splash. Do not infer that differently named aliases use identical weights or only differ by engine. Source Format-column label: Splash. Exact measured quantization: unknown. Measured weight format: unknown; source label is not a confirmed weight container. Application: LM Studio. Backend: Splash. Exact app/backend versions are unknown. Do not turn QAT, MLX or Splash labels into an assumed Q4 quantization. An official Qwen3.8 Splash package is publicly available with a packed 4-bit target and DFlash 2 draft. This availability does not establish the package, bit width, revision or hash actually used by the author. Verified public related package/collection: https://huggingface.co/incoai/Qwen3.8-27B-Splash/tree/main Measured Hugging Face repository, filename, revision and hash are unrecorded. Public related-package links above prove availability only. The source identifies Splash speculative decoding with DFlash2 drafts. This is not evidence of a native MTP head setting. Exact draft checkpoint/version, target quantization and package manifest are unknown. The three Splash aliases are preserved as separate author-measured configurations, not pooled or assumed to share identical weights. 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 15.3 tokens/s: Run 0: 14.3 tokens/s; 516 output count; 515 chunks; TTFT 0.42 s. Run 1: 16.1 tokens/s; 575 output count; 574 chunks; TTFT 0.29 s. Run 2: 15.3 tokens/s; 600 output count; 599 chunks; TTFT 0.27 s. German prose, bead-based immunoassay (beads_de), median 17.2 tokens/s: Run 0: 17.6 tokens/s; 538 output count; 537 chunks; TTFT 2.69 s. Run 1: 17 tokens/s; 502 output count; 501 chunks; TTFT 0.29 s. Run 2: 17.2 tokens/s; 562 output count; 561 chunks; TTFT 0.27 s. English prose, Ruhr economy (prosa_en), median 20.2 tokens/s: Run 0: 20.6 tokens/s; 461 output count; 460 chunks; TTFT 1.78 s. Run 1: 20.2 tokens/s; 479 output count; 478 chunks; TTFT 0.28 s. Run 2: 20 tokens/s; 476 output count; 475 chunks; TTFT 0.27 s. Python ISO-8601 duration parser (code_py), median 45.2 tokens/s: Run 0: 40.1 tokens/s; 600 output count; 600 chunks; TTFT 0.4 s. Run 1: 45.2 tokens/s; 600 output count; 600 chunks; TTFT 2.25 s. Run 2: 52.1 tokens/s; 571 output count; 570 chunks; TTFT 1.92 s. JSON antigen/disease extraction (json_extr), median 67.7 tokens/s: Run 0: 68.2 tokens/s; 87 output count; 86 chunks; TTFT 0.48 s. Run 1: 67.7 tokens/s; 87 output count; 86 chunks; TTFT 0.27 s. Run 2: 67.6 tokens/s; 87 output count; 86 chunks; TTFT 0.27 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.