How Darkbloom decides which Mac gets a request
When someone sends a request to a model, Darkbloom’s coordinator picks one Mac from those serving that model. Knowing how it picks explains why two Macs running the same model can earn very different amounts. This is based on Darkbloom’s open-source routing docs.
Fastest estimated answer wins
- For each candidate Mac the coordinator estimates how long the request would take: whether the model is warm, the queue, expected prompt-processing and generation time, and penalties for memory, GPU or thermal pressure and recent capacity rejections.
- The lowest estimate wins. When Macs are within 3 seconds of each other, the one with the shortest queue wins, then it’s random.
- A cold or unknown slot is charged a large time penalty, so a Mac that just loaded a model competes poorly until it has served a few requests.
No reputation score
Darkbloom’s routing docs say it plainly: “There is no composite reputation score.” Past uptime and job history don’t earn you a better position. Speed relative to the other warm Macs does.
Gates before the race
- A Mac only receives models it advertises.
- Dedicated models: by default Gemma 4 is a dedicated family. Public Gemma requests only go to Macs whose whole advertised list is Gemma 4.
- Macs in cooldown, failing health checks or ejected for errors are skipped.
The coordinator also moves models around
A warm-pool controller checks every 10 seconds how many warm Macs each model needs and asks idle Macs to load models they advertise and already have on disk. A Mac that advertises several models can be switched between them by the network itself.
What the network looks like
In public network stats from September 27, 2026 (about 1,140 providers), Macs with 96 GB or more were 29% of providers and generated 60% of tokens. Macs with 32 GB or less were 18% of providers and generated about 1%. The network had far more capacity than demand: under 6% of it was in use.
What this means for you
- Keep the model warm and the Mac idle apart from Darkbloom. Queue and pressure penalties cost you requests.
- If you run Gemma, advertise only Gemma.
- On a slower chip, a model that fewer fast Macs serve can earn more than the popular one.
Sources
- Darkbloom routing architecture: github.com/Layr-Labs/d-inference/blob/master/docs/architecture/routing.md
- Configuration reference (dedicated models): github.com/Layr-Labs/d-inference/blob/master/docs/reference/configuration.md
- Network figures: Darkbloom’s public /v1/stats endpoint, three snapshots on September 27, 2026.
How Bloomkeeper helps
Bloomkeeper’s network view shows demand per model and how many Macs serve it, and the Manager uses what Macs with your chip and memory actually earn, not list prices, to decide what to run.
Questions
Does Darkbloom have a reputation score for providers?
No. Darkbloom’s routing docs state there is no composite reputation score; each request goes to the Mac with the lowest estimated response time among those serving the model.
Why does my Mac get fewer requests than others on the same model?
Routing favors the fastest expected answer, so faster chips, warm models and empty queues win. Within 3 seconds of each other, the shortest queue wins, then it is random.
Why am I not getting Gemma requests?
By default Gemma 4 is a dedicated model: public Gemma requests only go to Macs that advertise nothing but Gemma 4. Remove other models from your advertised list.
Related
- How Darkbloom pays Mac providers: token earnings and base rewards
- Which Darkbloom model should I run on my Mac?
- Why is my Mac online but not getting Darkbloom jobs?
Updated 2026-09-28. Still stuck? Ask in #bloom-dash-optimizer on the Darkbloom Slack or contact us. Bloomkeeper is independent and not affiliated with Darkbloom.