Darkbloom.dev (Private inference on idle Macs)

[fleet ingot]

tldr;

• Open-AI compatible API

• Cheap inference: 50% off what OpenRouter charges: https://darkbloom.dev/

• These are the current models: https://console.darkbloom.dev/models

• If you have a M-series Mac you can plug it into the network and others can use your unused compute for model requests (it's like open router for macs just sitting around). https://console.darkbloom.dev/providers

• You probably won't make more than the electricity costs, but it might make you feel good to help others out with your unused Mac that is just sitting there. https://console.darkbloom.dev/earn

• Anyone who can't run these models locally, it's a really good option for them.

I've been running a provider on my M3 Ultra for the last two days. The vast majority of the requests are to Gemma-4 https://console.darkbloom.dev/stats.

P.S. If you really want to know what earning is like at the moment, I have attached my last 100 requests to my Mac Studio M3 Ultra 512 GB provider on there. You can have your AI agent extrapolate out what you might make in a month.

Update: Stripe Connect is now wired up so you can buy / earn! https://console.darkbloom.dev/billing

[deep gust]

Does it auto load and unload models or do you pick one or two to run? No quantization?

[near sparrow]

I wouldn't trust this inference, there's nothing that actually guarantees providers do the inference (either at all or on the requested model).

[fleet ingot]

You pick the ones you want to run but you have to have enough RAM to run all (at the same time) the ones you pick. It will unload them automatically after an hour if there are no requests to them to save on power. Then it loads them again automatically if they are needed.

[fleet ingot]

That’s your call, but this actually has a lot more guarantees than say OpenRouter. The entire premise is verified compute. Go look at EiganLabs the company behind it.

If any of the code is changed, or a single bit is different then the attestation fails.So unlike OpenRouter you know EXACTLY the model and the server code and the quantization that is being provided to you.