← Darkbloom help

Which Darkbloom model should I run on my Mac?

Updated

Short answer: run Gemma 4 26B (QAT 4-bit) on its own as your steady earner if your Mac has 32 GB or more, keep gpt-oss-20b as the fallback, and treat a large Qwen (Qwen3.5 or Qwen3.6 35B, 36 GB and up) as an alternative, not something to chase: switching to follow demand spikes has usually paid less than staying on Gemma (see why BloomGauge rarely switches models). On a 64 GB Mac every current model fits. What a model pays depends on how many people are asking for it right now and how many other Macs already serve it. That changes through the day, so there is rarely one right answer for long. These rules of thumb come from weeks of Darkbloom history measured by BloomGauge.

Three kinds of model

Gemma alongside other models

Darkbloom’s open-source coordinator can make Gemma 4 a dedicated model, so that public Gemma requests only go to Macs that advertise nothing else. Darkbloom staff said in August 2026 that this setting is turned off on the live network, and providers who serve several models report getting Gemma requests too. Running one model is still usually best, because models loaded together share memory and memory bandwidth. In Darkbloom’s public stats (polled every 10 minutes, Sep 24–Oct 1, 2026), Macs offering Gemma plus other models got only about 0.3× the Gemma requests per hour of Gemma-only Macs, mostly because they spent half their time serving the other model, and earned about half as much per hour overall.

Which models fit, by memory

Free memory needed when the model loads, and the smallest Mac where it loads with nothing else running, from Darkbloom’s hardware requirements (September 2026). Re-check with darkbloom models catalog, because the catalog changes. Fitting is not the same as earning: in Darkbloom’s public stats (early October 2026), 24 GB Macs almost never had a model loaded, and 32–36 GB Macs usually earn very little. New providers need 48 GB or more.

ModelFree memory at loadSmallest Mac
gpt-oss-20b18.0 GiB24 GB (tight)
Gemma 4 26B (QAT 4-bit)23.9 GiB32 GB
Qwen3-VL 30B A3B27.0 GiB32 GB (tight)
Qwen3.5 35B A3B29.9 GiB36 GB
Qwen3.6 35B A3B30.3 GiB36 GB
Gemma 4 26B (8-bit)37.8 GiB48 GB

What to run, by memory

Memory decides what fits

The model has to fit in your Mac’s unified memory next to everything else. macOS keeps the last model’s files in its file cache and Darkbloom counts that as in use, so a large model can refuse to load right after a switch even though it would fit. Clearing the file cache (sudo purge) before loading fixes that.

Measure it on your own Mac

Pay depends on your chip, memory and how many other providers serve the same model, so numbers from someone else’s Mac only go so far. Run a model for a few hours at different times of day before you decide it doesn’t pay.

How BloomGauge helps

With Manager on, BloomGauge runs Recommended, the model expected to pay most on your Mac, and moves only when another beats it by 25% after switch costs for 15 minutes, at most twice an hour, or holds one model you pick. At most once a day it may try a model with no real pay yet for an hour.

Download BloomGauge, free for Mac

Questions

What is the best Darkbloom model to run?

It changes with demand and with your Mac. Gemma has been the steadiest earner on Macs with 36 GB or more (usually best run on its own), large Qwen models can pay more during demand spikes (though switching to chase them has usually lost), and gpt-oss is a dependable fallback when the others are quiet.

Which Darkbloom model should I run on a 64 GB Mac?

Every current catalog model fits on 64 GB. Gemma 4 26B, run on its own, has been the steadiest earner; large Qwen models can pay more during spikes, but chasing them has usually paid less than holding Gemma; gpt-oss-20b is the fallback. Compare what each actually pays on your Mac over a few days, or let a manager such as BloomGauge do it.

Why won’t a large Darkbloom model load on my Mac?

Often because macOS is still holding the previous model’s files in its file cache, which Darkbloom counts as used memory. Clearing the cache with sudo purge before loading usually fixes it; otherwise the model is too large for your Mac’s memory.

Related

Updated 2026-10-03. Still stuck? Ask in #bloomgauge on the Darkbloom Slack (join it here, free) or contact us. BloomGauge is independent and not affiliated with Darkbloom.