Aleph Alpha released Kolibri 1, an open-weight 78 billion parameter model for German and English, under Apache 2.0

A German team released a 78 billion parameter open-weight language model today, written to run on a customer's own servers under European law. Aleph Alpha published Kolibri 1 on 3 October 2026 under Apache 2.0, with the full weights on Hugging Face, and the company says it was trained from scratch on infrastructure in Germany and Finland.

Kolibri is a mixture-of-experts model with 78.1 billion parameters in total and 3.46 billion active for each token, which Aleph Alpha says lets it serve 18 concurrent 256,000-token requests on two H100 GPUs where a dense 123 billion parameter model handles three. The model has 50 layers, 40 of them reading the last 512 tokens only and every fifth layer reading the full context. It supports 262,144 tokens of context natively and was validated at 1,048,576. There are four reasoning effort levels: none, low, medium and high.

The company trained a bilingual tokenizer with 128,000 entries using a method it calls UniBPE. On its own measurement, the tokenizer needs 11.2 percent fewer tokens for German text than the one GPT-5 uses. An independent test from Tejas Kumar, an AI engineer at IBM, ran six tokenizers over the German Basic Law and found Kolibri's used 15 percent fewer tokens than GPT-5's on legal German, and tied with it on the official English translation. Aleph Alpha trained on 24 trillion tokens over three stages on 768 NVIDIA B200 GPUs. German made up 21.3 percent of pre-training, 4.3 trillion tokens, 80 percent of which Aleph Alpha curated or rephrased from German web text.

Aleph Alpha's own benchmark table says Kolibri scores 96.9 on AIME 2025 math, 84.3 on GPQA Diamond knowledge, and 85.9 on LiveCodeBench v6, matching or beating models with up to four times its active parameter count. These are the company's own measurements. Public scores for the model from independent evaluators were not available on release day. The company signed the European Union's General-Purpose AI Code of Practice and says the model was built with the EU AI Act, that Code of Practice and the GDPR in mind.

Teams in European public administration, aerospace, manufacturing and industrial groups that have been blocked from sending documents to US-hosted APIs now have an open-weight model that matches models four times its active size on math, code and knowledge, written for German and licensed to modify. Running it has a cost: the model holds all 78.1 billion parameters in memory even though only 3.46 billion do work for each token, so an on-premise deployment needs 78 GB of 8-bit weights plus the overhead of each worker. Teams willing to pay that cost get a model that stays on their own hardware, with the full weights, the technical report and the licence to change it.

Source

• Aleph Alpha: Kolibri Has Landed: A Sovereign Open-Weight Model
• Model page: Aleph-Alpha/Kolibri-1 on Hugging Face
• Third-party technical analysis: Tejas Kumar on how Kolibri works