Aleph Alpha released Kolibri’s weights on 3 October, followed by a newsroom announcement on 5 October. The German-English language model supports document-grounded answers and tool use, with deployment on infrastructure controlled by the customer.

Based on 4 primary sources · See sources

Open weights, with a context caveat

The model card lists 78 billion total parameters and 3.46 billion active per token, under Apache 2.0. It supports roughly one million context tokens, but recommends no more than 262,144 for serving efficiency and complex tasks.

From the source · Aleph Alpha · @Aleph__Alpha ·

Aleph Alpha releases Kolibri weights for deployment on customers’ own hardware.

View the original post on X ↗

A deployment option for technical teams

The card puts the FP8 weight footprint at about 78 GB and lists server GPUs among minimum configurations. Open weights do not mean cost-free hosting or an immediately usable marketing application. Aleph Alpha says the model was developed in Germany and trained in Germany and Finland.

For marketers

For teams building German-language assistants around internal product and brand documents, local deployment offers a different way to control infrastructure. That is an architectural option, not evidence of better campaign results or automatic compliance.

Your next move

Evaluate answers against your own German and English product questions. Compare hosting, retrieval quality and operations effort alongside the licence.

Sources

Written with AI assistance from the sources above. The marketing implications are our interpretation. Our editorial approach.