Local LLM Benchmarks speed dataset (Donato Capitella, results.json)

Machine-readable export behind local-llm-benchmarks.dev/speed. 114 completed runs, format_version 2, runner local-llm-speed v1, protocol deterministic-prime-extend-completions-v1, measurement_layer api-endpoint over an OpenAI-completions endpoint. Each point is the mean of 2 repeats after 1 warmup. Concurrency is per point, not per run: the same run carries concurrency 1, 4, 8 and 16. Only the concurrency-1 points are per-request; the higher ones are aggregate throughput of the server. Depth is the primed KV prefix; at depth 0 the whole prompt is computed (actual_cached_depth_tokens 0, recomputed_history_tokens 0). Prefill is effective follow-up prompt processing: actual uncached tokens divided by client-observed time to first output. Decode is client-observed generation throughput: subsequent output tokens divided by the first-to-last token interval, so it excludes TTFT. Platforms present: AMD Ryzen AI Max+ 395 (Strix Halo, 63 runs), AMD Radeon AI PRO R9700 (36), and an unidentified gorgon-halo host (15). Engine versions are only populated on the newer runs; the older ones name the runtime only (ROCm 10.0 or Vulkan RADV). Accessed 2026-10-17, newest run completed 2026-10-09.