Lumen
From Wikipedia, the free encyclopedia
 | |
| Type | Private company |
|---|---|
| Industry | Artificial intelligence |
| Founded | 2018 |
| Headquarters | Munich |
| Main data centre | Frankfurt |
| Products | Lumi |
Lumen is an artificial intelligence company in the European Federation, based in Munich. It develops the Lumi family of language models, which it offers to the public as cloud services and as open-weight releases. Its development has been funded by Federation institutions, and it trains and serves its models in its own data centres, the largest of them in Frankfurt. Lumi is widely used in the Federation, often as the preinstalled or free option.
History
Lumen was founded in Munich in 2018 by researchers from the city's Technical University, with which it still works closely. Its offices are in the Maxvorstadt, near the universities. Federation research institutions have funded it from its first years.
Its early models scored below those of the leading American laboratories on benchmarks. The American laboratories scaled their models faster, financed by successive rounds of venture capital, while Lumen favoured efficient training over increases in scale. Since 2026 its largest models have scored close to the American ones.
The first stage of its Frankfurt data centre opened in 2021, and it has been extended twice since. Since 2025 computing capacity has been scarce across the world, and the federal government classes the data centre as critical infrastructure: it has priority on the power grid and is guarded by H&A Security.
Products
Lumen sells four Lumi models through its cloud service, in ascending order of size: Lumi Hymn, Lumi Lux, Lumi Nocturne and Lumi Requiem. Apart from Hymn, they are named after degrees of light, with darker names for larger models. Hymn and Lux are free to use within usage limits; Nocturne and Requiem require a subscription or API access. None of the four is published with open weights.
Lumen also publishes open-weight models, named by line rather than by version number: Vesper, a mixture-of-experts model with 26 billion parameters, and the much smaller Candela. They 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, and they run on the user's own hardware.
Lumen also keeps internal models that are not offered to the public. The best known is Parsec, which H&A uses to train its own models, among them the Beast and its Dispatcher decoders.
Lumen licenses a signal reader from H&A, which Lumi models can call to interpret intercepted Dispatcher signals.
Lumi models are among those run by agent swarms, alongside American and other European models.
Privacy
Lumen's services are subject to the Federation's data-protection law, which allows users' conversations to be used for training only with their consent, and requires the data of Federation users to be stored and processed within the Federation. Lumen trains its models in the Federation.
Reception
Federation officials have promoted Lumi as a gain for European sovereignty, while continuing to use American models themselves. The Federation's own institutions use models supplied by Mistral.
External links
- Lumi Candela on Hugging Face. The current Candela release.