Kolibri

Kolibri is a specialized sovereign large language model from Aleph Alpha, trained from scratch in Germany. It is a 78B-parameter Mixture-of-Experts model with 3B active parameters per token, built for German and English. It is listed as Generally Available, with open weights under the Apache 2.0 license on Hugging Face.

What It Is

Kolibri is a general-purpose LLM aimed at assistants and agentic workflows that need multi-step reasoning, structured extraction, retrieval-augmented generation and tool calling. Aleph Alpha describes it as specialized in German, using in-house curated German-language data, and as aligned with European legal and regulatory standards including the EU AI Act.

Model Specifications

The model page lists a context length of 1,048,576 tokens, with 256k tokens recommended for serving efficiency and complex tasks. It uses bfloat16 precision, has a controllable reasoning mode and supports tool calling. The memory footprint is about 78 GB with FP8 weights. The minimum hardware is 2x A100 80 GB, 2x H100 SXM5, 1x H200, 1x B200 or 1x B300. Knowledge cutoff is listed as Jun 18, 2026 for both English and German.

Performance Claims

Aleph Alpha publishes benchmark comparisons against Qwen3.6 35B-A3B, Mistral Small 4 and Nemotron 3 Super. They cover tool and agent use (Tau2/Tau3-Bench, BFCL v4), German language skills, and reasoning tasks. The vendor also states that Kolibri performs comparably to larger models while generating more text per GPU.

Access and Deployment

Users can download the weights from Hugging Face or contact Aleph Alpha sales for enterprise deployment and specialization. A model card, a sufficiently detailed summary, a technical report and a release blog post are provided.