The dataset

41103 measured records, retrieved 2026-10-22. Download and diff: the data is the product, not the site. Reference estimates are kept apart from records (see below).

Schema

fieldmeaning
idstable unique slug for the record
modelmodel name as published by the source
paramsparameter size, as published (e.g. 8B, 3.8B)
quantquantization (e.g. Q4_K_M)
hardwarechip/hardware identifier (e.g. M5 Max, RTX 4090)
ram_gbmemory, GB (when the source states it)
backendinference backend/engine (llama.cpp, MLX, Ollama, vLLM, …)
ctxcontext depth the benchmark actually ran at (when stated; null when the harness did not fix a context)
batchbatch size / concurrency (when stated)
tpsgeneration tokens/second
pp_tpsprompt-processing tokens/second (when reported)
pp_tokensprompt token length of the prompt test (e.g. 512 for pp512)
tg_tokensgeneration token length of the decode test (e.g. 128 for tg128)
scoperecord scope: empty = in scope; cluster = multi-node/cluster run, reference area only
ttft_stime to first token, seconds (when reported)
power_wpower draw, watts (when reported)
flagscomputed flags: contradiction / outlier (see schema.md rules 4-5)
dateoriginal measurement date, as published
provenancesourced | community | estimated (see schema)
source_urlURL the number was retrieved from
source_namename of the source dataset/page
retrieveddate this record was retrieved
quoteexact values as they appear in the source (comma-joined row)
notescollector notes

No record without provenance

Every row carries source_url, retrieved, and quote (the exact values as published). estimated rows (the source's own model-based estimates) are reference values, not records: they live in data/reference/estimates.json, are excluded from record counts, and are only shown where explicitly labeled (lookup toggle, cross-source checks).

Rules and field semantics: data/schema.md.