Darkbloom 0.9.10: live model switching and on-demand loading
Darkbloom 0.9.10 changed how the provider starts and switches models. Check yours with darkbloom --version. Darkbloom’s changelog still lists 0.9.10 and later as release candidates, but they are running on provider Macs.
What changed
- darkbloom switch replaces the model a running provider serves. It drains first (accepted requests finish, up to 600 seconds by default) and keeps the same connection to the network, with no restart, reconnect or re-attestation.
- Routing to the Mac resumes only after the provider confirms the new model is ready.
- Selected models preload on every provider start, whatever the idle setting.
- darkbloom models list shows only enabled models. Add --all to see everything on disk.
Gotcha: advertised but not loaded
A model can be advertised and still cold. A cold model gets no requests and, if nothing else is loaded, no base reward. On our test Mac a failed load left a larger model advertised but cold for over an hour and a half, earning nothing, until the previous model was started again by hand.
Gotcha: config and running model disagree
If the provider config’s enabled models and the model you start with disagree, a restart can come back on the wrong model. After switching, run darkbloom status and check the model that is actually loaded.
Checklist after any switch
- darkbloom status: the new model is loaded and serving.
- Give it a few minutes: a freshly loaded model competes poorly until it has served some requests.
- If nothing arrives for 10–20 minutes, restore the previous model.
Sources
- Darkbloom changelog: github.com/Layr-Labs/d-inference/blob/master/CHANGELOG.md
- darkbloom switch --help and darkbloom idle --help on 0.9.10.
How Bloomkeeper helps
Bloomkeeper checks that a switched model is warm and serving before calling it ready, restores the previous model if a load fails, and restarts a provider it finds drained.
Questions
What does darkbloom switch do?
Added in 0.9.10, it swaps the model a running provider serves after a graceful drain, on the same network session, without restarting the process, reconnecting or re-attesting.
Why is my Darkbloom model advertised but not getting requests?
It may be advertised but not loaded (cold). Cold models get no requests, and with nothing loaded base rewards pause. Run darkbloom status to see what is actually loaded.
Related
- Darkbloom provider stuck draining after a restart
- Why is my Mac online but not getting Darkbloom jobs?
- How Darkbloom pays Mac providers: token earnings and base rewards
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.