Running Ollama or another local AI server next to Darkbloom
Updated
Short answer: the Darkbloom provider doesn’t listen on Ollama’s port, so the two don’t fight over 11434. They compete for the same unified memory and GPU, and darkbloom doctor warns about that as “competing inference”. Quit Ollama, or unload its models, while your Mac serves Darkbloom. A real port clash only happens with Darkbloom’s optional local endpoint, which uses port 8000 by default. Everything below comes from Darkbloom’s docs and code and from Ollama’s own FAQ (links at the end).
What conflicts, and what doesn’t
- Network: the provider only connects out, to api.darkbloom.dev on port 443. It needs no inbound port, so a program listening on 11434 can’t block its traffic.
- Memory and GPU: Ollama, llama-server, mlx_lm.server, vLLM and text-generation-launcher all load models into the same unified memory and use the same GPU as Darkbloom. darkbloom doctor looks for these, and for anything listening on 11434, and warns that they “can reduce usable RAM / deroute paid work”.
- Less free memory means a smaller token budget for your Mac, a model that won’t load, or slower answers. Each of those means fewer paid requests.
- Ports only matter for Darkbloom’s own local endpoint (darkbloom start --local or --local-endpoint), which listens on 127.0.0.1:8000 unless you choose another port.
Check with darkbloom doctor
- Run darkbloom doctor and find the competing inference line. ✓ means nothing was found.
- A ⚠ names what it found, for example port 11434 LISTEN (Ollama default) or processes: ollama.
- Quit the other app, then run darkbloom doctor again.
See what is listening yourself
- lsof -nP -iTCP -sTCP:LISTEN lists every program listening on a TCP port, with its name and port.
- lsof -nP -iTCP:11434 -sTCP:LISTEN checks Ollama’s default port. It is the same check darkbloom doctor runs.
- lsof -nP -iTCP:8000 -sTCP:LISTEN checks the port Darkbloom’s local endpoint uses by default.
- ollama ps shows which models Ollama has in memory right now.
Free the memory
- Quit Ollama from its menu bar icon. That stops its server and frees its models.
- Or keep Ollama and unload its models with ollama stop <model>. By default Ollama keeps a model in memory for 5 minutes after the last request.
- Run darkbloom status and check that your Darkbloom model is loaded and not listed as Cold load blocked (memory).
“Local server failed to bind” (port 8000)
- darkbloom start --local stops with “Local server failed to bind <addr>:<port> within 5s” when another program holds the port. Choose another port with --port, or stop that program.
- With --local-endpoint, a busy port doesn’t stop the provider. It logs “Local OpenAI endpoint did NOT bind … (port already in use?)” and keeps serving the network. Restart with a free port if you want the local endpoint.
- Example: darkbloom start --local-endpoint --port 8080
Sources
- Doctor checks and start errors: github.com/Layr-Labs/d-inference/blob/master/docs/provider/troubleshooting.md
- Local endpoint and ports: github.com/Layr-Labs/d-inference/blob/master/docs/provider/direct-mode.md
- Network needs (outbound only): github.com/Layr-Labs/d-inference/blob/master/docs/provider/hardware-requirements.md
- Competing inference check: github.com/Layr-Labs/d-inference/blob/master/provider-swift/Sources/darkbloom/Diagnostics/CompetingInferenceDiagnostics.swift
- Ollama FAQ (port 11434, ollama ps, ollama stop, 5-minute default): github.com/ollama/ollama/blob/main/docs/faq.mdx
How BloomGauge helps
BloomGauge serves its own dashboard only on 127.0.0.1:8765 (and 8766 for the phone view), so it doesn’t take Darkbloom’s port 8000 or Ollama’s 11434. It shows CPU, GPU, memory and temperature readings next to your earnings.
Questions
Can I run Ollama and Darkbloom on the same Mac?
Yes, but not at the same time if you want steady Darkbloom earnings. They don’t share a port, but both load models into the same unified memory and use the same GPU. Quit Ollama or unload its models with ollama stop while the Mac serves Darkbloom.
Does Ollama on port 11434 block Darkbloom?
No. The Darkbloom provider only connects out to api.darkbloom.dev on port 443 and listens on no port by default. darkbloom doctor checks 11434 only to spot Ollama, because Ollama competes for memory and GPU time.
What does “competing inference” mean in darkbloom doctor?
Another AI server, such as Ollama, llama-server, mlx_lm.server or vLLM, is running on the Mac or listening on Ollama’s port 11434. It can reduce the memory Darkbloom can use and cost you paid work. Quit it and run darkbloom doctor again.
Related
- Darkbloom model won’t load: not enough memory
- Can you run several Darkbloom models at once?
- Darkbloom “machine_busy” or “your machine is at capacity” while the Mac is idle
- Why is my Mac online but not getting Darkbloom jobs?
Updated 2026-09-29. Still stuck? Ask in #bloomgauge on the Darkbloom Slack or contact us. BloomGauge is independent and not affiliated with Darkbloom.