Kolibri-1: Aleph Alpha releases a German-English reasoning model with 78 billion parameters under the Apache 2.0 licence

Aleph Alpha’s first open-source model is a mixture-of-experts model with 78 billion parameters – 3.46 billion of which are active per token – focused on German and English, with a context validated up to one million tokens.

Aleph Alpha has uploaded Kolibri-1, an open-weight reasoning model released under the Apache 2.0 license and dated October 3, 2026, to Hugging Face. The architecture is a mixture-of-experts of 78 billion parameters, of which only 3.46 billion are requested for each token, which limits the computational cost per use. The model handles an explicit reasoning mode and tool calling, and the publisher announces a validated quality of up to 1,048,576 context tokens, while recommending to stay under 262,144 tokens in production.

Two languages rather than fifty

The choice is assumed: to focus on depth in German and English rather than on the coverage of many languages. The pre-training corpus has 20,000 billion tokens (about 62.5% English, 23.9% German and 13.6% code), and a tokenizer was designed for the German morphology. The training mobilized 768 Nvidia B200 GPUs for 21 days for the main phase alone; the total consumption is estimated at about 950 MWh.

Aleph Alpha, a signatory of the European Code of Practice on AI for general use, is intended for Kolibri as assistants and agentic chains where a human reads the results before acting. In FP8, the model occupies about 78 GB of memory: a single B200 or H200, or two H100s, are enough to serve it.