Why BloomGauge rarely switches models (and when it does)

Updated October 2, 2026

Short answer: chasing live demand has paid less than staying put. On our test Mac, even perfect hindsight (the best model every hour) would have added only about 6% over holding Gemma, and every rule that has to decide in real time came out about break-even or worse. So the Manager holds one home model and moves only on strong, sustained evidence. The numbers below come from one Mac with real pay plus public Darkbloom network data for other hardware, so treat them as a first read, not a law.

“That model is quiet right now, why not switch?”

Because by the time a switch pays off, the moment has usually passed. Busy stretches on a model typically last 20 to 30 minutes, and once one is 30 minutes in, only 17–50% carry on for another hour. A switch also costs time: about 7 minutes of paid work, plus about one 5-minute base-reward period while the new model loads. Even with a perfect pay estimate, the best real-time switching rule on our test Mac earned about the same as just holding Gemma.

What the replays found

On our test Mac, perfect hindsight (the best model every hour, Sep 11–30) would have added only about 6% over holding Gemma. On big Macs that ceiling is larger, but no rule we tested captured it.
In a 20-day replay, 38 of 43 spike trips paid less than staying on Gemma.
A first replay on about 2 days of public network data lost on every hardware type tested, and lost more on big Macs, whose Gemma home pays more and so costs more to leave.
BloomGauge’s earlier pay estimate for non-Gemma models ran 46–138% too high on days it hadn’t seen. It is now corrected against each Mac’s own results.
On our test Mac, the old demand-chasing optimizer earned about 16% less than holding Gemma. The current Manager comes within about 3% (early days).
In public network data, the home model picked on the first half of the period was still the best on the second half for 24 of 29 hardware types (list prices, about 3 days).

Switching has hidden costs

Loading a new model takes the Mac out of the running for paid work for a minute or two, and pay takes a while to get back to normal. All told, about 7 minutes of paid work plus about one 5-minute base-reward period per switch.
The first paid job can be slow to arrive: 10 of 61 Gemma restarts on our test Mac (16%) waited more than 10 minutes for one.
Those costs are small on their own. They add up when a spike ends before the new model has earned them back.

How the Manager picks a home

The home model is your pick if you pinned one; otherwise it is the model that has paid best on this Mac over the last 30 days. Until a Mac has 3 days of its own history, BloomGauge starts from what pays best on Macs with the same chip and memory. On almost every Mac with 36 GB or more that has been Gemma 4 26B; on 32 GB Macs it has been gpt-oss-20b.

When it does move

With “Switch to better models when network evidence is strong” on (the default), it leaves home only when at least five Macs like yours have clearly earned more on another model for two hours in a row. At most 3 moves a day.
Before it changes your home model it tells you first, with a one-tap Keep current button.
It comes back when the evidence fades or the other model pays less than home would, and it turns these moves off on its own if they haven’t clearly paid off.
It runs no blind trials of other models.

What it does instead: keeps the home model earning

Most lost earnings come from a Mac that stops working, not from the wrong model. Keep this Mac earning watches for that and runs stall recovery: a test request first, then a restart, within limits. In one public Darkbloom issue, 43 of 44 affected Macs recovered within 1–3 minutes of a restart. Recovery also waits out the short network-wide dip that often follows a Darkbloom update (in 9 of 14 recent updates, routable Macs dropped 10% or more within 2 hours; ordinary 2-hour windows show a similar dip about 13% of the time, so it’s a pattern, not proof). Keeping a model loaded matters too: in public network snapshots, about 30% of online Macs had no model loaded, which earns no base reward.

If you want to chase demand anyway

You can. Turn the evidence switch off and switch models yourself from model controls, or pin a different home. BloomGauge’s demand alerts tell you when a model gets unusually busy. Just know that, so far, the data says those trips more often lose than win.

What this data can’t tell you yet

Real pay comes from one Mac; other hardware comes from public request data priced at list rates, over a few days. Big Macs show a larger hindsight ceiling, so a per-hardware, dollar-based switching rule may pay there later. A longer public data collection is running, and this article will be updated as it lands.

How BloomGauge helps

BloomGauge’s Manager holds one home model (usually Gemma 4 26B on 36 GB and up), moves only when Macs like yours have clearly earned more elsewhere for two hours, tells you first with Keep current, and spends its effort keeping the home model loaded and recovering stalls.

Download BloomGauge, free for Mac

Questions

Why doesn’t BloomGauge switch to whichever Darkbloom model is busy right now?

Busy stretches usually end within 20–30 minutes, before a switch pays back its cost. In replays, 38 of 43 spike trips paid less than staying on Gemma, and even perfect hindsight added only about 6% on our test Mac.

Does BloomGauge ever switch Darkbloom models?

Yes, when at least five Macs like yours have clearly earned more on another model for two hours in a row, at most 3 times a day. It tells you first and offers Keep current.

Can I switch models myself?

Yes. Use model controls, pin a different home model, or turn off “Switch to better models when network evidence is strong”.

How much does switching a Darkbloom model cost?

About 7 minutes of paid work plus about one 5-minute base-reward period per switch on our test Mac, and more when the first paid job is slow to arrive.

Updated 2026-10-02. Still stuck? Ask in #bloomgauge on the Darkbloom Slack or contact us. BloomGauge is independent and not affiliated with Darkbloom.