how paid frontier LLM burn dropped 91% with a 24GB M4 mac mini
paid frontier burn was $42.50 a day. after routine work moved onto a standard M4 mac mini with 24GB, it is $3.90 a day. that is a 91% cut, and the box has processed 40.36M local tokens since.
what moved local
the cheap, high-volume steps: reading context, sorting what needs doing, first drafts and checks. the paid models only see the hard steps, and every handoff is counted. there were 64 escalations in 12 days.
what keeps it fast enough
- a 35B-total mixture-of-experts with about 3B active parameters, so each token reads a small slice of the weights
- a 2,048-token prefix pinned in KV cache at an 88.4% hit rate
- every background job runs in 20s slices under a hard timeout, after two RAM freezes
what keeps it honest
- a deterministic blocklist runs before any model judge
- the judge fails closed: bad JSON, empty output or an unreachable model all score 0
- 7 specialized agents share the box, and nothing outward happens without a human approving it
the full speed write-up is in 41 tok/s from a 35B model. the full cost math is in i finally priced the work.
want the setup behind this? browse the open-source repos, or book an automation audit.