the write-ups
benchmarks, flags, cost math and the gates that keep agent work honest. every figure traces back to a log on the box.
i finally priced the work running on my mini
16 days of local model traffic on a base M4 mac mini with 24GB, priced at API list rates: $165.61, or $310.52 per 30 days.
read it →41 tok/s from a 35B model on a base M4 mac mini with 24GB
the exact llama.cpp launch line, the benchmarks and the draft settings that took a base M4 24GB from 22.2 to a 41.3 tok/s peak.
read it →the draft looked fine until i checked the claim
nine quality-gate rules for agent drafts at work: blocklists before judges, written rubrics, fail-closed scoring and holding claims with no source.
read it →how paid frontier LLM burn dropped 91% with a 24GB M4 mac mini
moving routine agent work onto a $599 base M4 mac mini with 24GB took paid API burn from $42.50 a day to $3.90 a day.
read it →daily findings
what the model server measured today: decode, prefill, draft acceptance, memory.
read the log →what broke and the rule kept
dated notes on failures, the wrong theory, and the measurement that settled it.
open the journal →get the build log
one short email when something new ships: a benchmark, a cost number, a fix. double opt-in, one-click unsubscribe.