Laptops sit idle most of the night. nocturne puts them to work: a small LLM API served by up to four Mac laptops that opens in the evening, closes in the morning and charges per request in sats.
Try it: lucashanson.dk/nocturne (opens in a new tab)
What you get
- OpenAI-compatible. Point any OpenAI client at it and change the base URL. Chat completions, streaming and thinking on or off.
- No account, no card. Get a key, top up over Lightning and pay only for the tokens you use.
- No prompt logging. Prompts and responses are not stored. Only token counts are kept, for billing.
- Night hours. Open roughly 17:30 to 08:00 Copenhagen time. Outside that window requests get
503with aRetry-Afterheader.
Honest limits
- It is a small model on laptops, not a frontier model in a datacenter.
- Capacity varies. Machines join and leave during the night: a laptop drops out when someone uses it or its battery drops below 20%. The service page shows how many are serving right now.
- Each request runs entirely on one machine, at about 18 to 23 tokens per second, with one or two requests per machine at a time.
- No uptime promise. It is an experiment.
How it works
A small coordinator takes requests and hands each one to an eligible worker. Workers only take jobs while idle, cool, and plugged in or above 20% battery, and they check in constantly. If a machine drops out mid-request, the request is retried once on another machine; streamed requests are not retried and stop with an error instead. Credits are reserved before a request runs and settled on the actual token count, so a failed request costs nothing. Prices are fixed per token in DKK and charged in sats at the current BTC rate. When Danish electricity gets expensive enough that a request would cost more in power than it earns, the service pauses instead of running at a loss.
Pricing, live status, the model name and the API reference are on the service page (opens in a new tab).