Lumi (language model)
From Wikipedia, the free encyclopedia
 The Lumi lantern, the mascot of the Lumi models | |
| Developer | Lumen |
|---|---|
| Type | Large language model |
| Cloud models | Hymn; Lux; Nocturne; Requiem |
| Open-weight models | Vesper; Candela |
Lumi is a family of large language models developed by Lumen, an artificial intelligence company in the European Federation. Lumen offers four Lumi models through its cloud service and publishes two lines of open-weight models, Vesper and Candela. Lumi is widely used in the Federation, often as the preinstalled or free option.
Models
Cloud models
Apart from Hymn, the cloud models are named after degrees of light, with darker names for larger models.
- Lumi Hymn is the smallest and cheapest. It is free to use within usage limits.
- Lumi Lux, named after the unit of illuminance, is the general-purpose model. It is free to use within usage limits.
- Lumi Nocturne is a large model, available by subscription or through the API.
- Lumi Requiem is Lumen's largest model, available by subscription or through the API.
All four are closed-weight and run only on Lumen's servers.
Open-weight models
Lumen's open-weight models are named by line, and the line name takes the place of a version number. The model files carry the family, line, size and tuning in their names.
Vesper, named after the evening star, is released as `Lumi-Vesper-26b-a4b-it`. It is a mixture-of-experts model with 26 billion parameters in total, of which about 4 billion are active for any one token, which lets it run on hardware too small for a dense model of the same size. The `it` marks the instruction-tuned release, as against the base checkpoint.
Candela, named after the unit of luminous intensity, is a much smaller model.
Both are released under the Lumen Open Weights Licence, which allows them to be used, run and modified freely under conditions required by the Federation's regulator. They run on the user's own hardware, and are trained mainly on English text.
As of 2028 Vesper is the open-weight model most used by businesses and municipalities in the Federation, which run it on their own hardware for work that cannot leave their premises.
Signal reading
Lumi models can call a licensed signal reader from H&A to interpret intercepted Dispatcher signals. The reader returns its interpretation with a confidence figure, which Lumi reports alongside it.
Privacy
Lumi's cloud services are subject to the Federation's data-protection law. Users' conversations are used for training only with their consent, and the data of Federation users is stored and processed within the Federation.
External links
- Lumi Candela on Hugging Face. The current Candela release.