Aleph Alpha Kolibri is a sovereign open-weight MoE model from Aleph Alpha, a German company that released it in Germany on October 3, 2026 under the Apache 2.0 license. It is built for public-sector and regulated-industry teams that must run capable language models on their own infrastructure under EU data rules.
Core Features
- Mixture-of-experts architecture with 78.1B total parameters but only 3.46B active per token, so it serves far cheaper than dense models of similar size.
- Native 262,144-token context, validated up to 1,048,576 (1M) tokens for long-document and agentic workflows.
- UniBPE tokenizer tuned for German compounds, cutting token use on legal and technical German text by up to 15% versus standard tokenizers.
- Four reasoning effort levels (none, low, medium, high) plus structured tool calling for agent pipelines.
- Open weights on Hugging Face, trained in Germany and Finland with the EU AI Act and GDPR in mind.
Use Cases
- EU public-sector and enterprise teams that cannot send data to third-party clouds.
- German and English legal, technical, or scientific document processing and drafting.
- Retrieval-augmented generation over internal knowledge bases.
- Long-context reasoning and multi-turn agentic workflows on owned hardware.
Pricing
Kolibri is free to download and self-host under Apache 2.0. Aleph Alpha does not sell it as a hosted API; you bring your own GPUs (roughly two 80GB cards or one H200/B200). Licensing cost is $0, with hardware and electricity as the only expenses.
Our Take
Best for European teams that need a sovereign, on-prem LLM with strong German support. Trade-off: it is German/English only and heavy to host at 78GB of FP8 weights, so smaller English-only shops should pick a lighter model. Browse more AI engine and model tools.
