Aleph Alpha has launched Kolibri, an open-weight AI model for public sector use, featuring advanced reasoning capabilities and extensive training data.
Aleph Alpha, a German tech firm, has launched Kolibri, an open-weight AI model designed specifically for governmental and regulated sectors. This announcement was made on October 3, coinciding with Germany's Unity Day, as highlighted on the developer's blog.
Kolibri employs a Mixture-of-Experts architecture and boasts 78.1 billion parameters, utilizing around 3.46 billion parameters for each token processed. It supports reasoning capabilities and external tool invocation, accommodating a context window of up to 1,048,576 tokens.
The complete model weights have been made available on Hugging Face under the Apache 2.0 license, with options in BF16 and FP8 formats.
Tailored for Critical Tasks
Described by Aleph Alpha as a "sovereign" AI model, Kolibri is intended for essential applications in public administration, industry, and aerospace. The development was carried out in Germany, utilizing local infrastructure while complying with EU and German regulations.
Organizations can implement Kolibri on their own systems, ensuring that sensitive internal data remains confidential and is not shared with third-party AI service providers. The model adheres to the European AI Act, the Code of Practice for General Purpose AI Models, and GDPR requirements, with safeguards to prevent it from providing responses in contexts lacking sufficient verification.
Extensive Training and Data Composition
Kolibri underwent extensive pre-training on 768 Nvidia B200 GPUs, lasting 21 days and incorporating 20 trillion tokens, each with a sequence length of 16,384. The model was subsequently fine-tuned with an additional 3.44 trillion tokens and adapted for extended context with 201 billion tokens, totaling approximately 23.64 trillion tokens across all training phases.
Notably, around 21.3% of the training dataset was in German, with English comprising about 62% and programming code making up 14%. Aleph Alpha emphasized that the German content was not largely derived from machine translations, with only about 6% being translated material.
A unique Anglo-German tokenizer was also created for Kolibri, accommodating the morphological intricacies of both languages, particularly the complex compound words in German. The model offers four reasoning modes: none, low, medium, and high.
Performance Comparisons
Aleph Alpha reported that Kolibri achieved a score of 96.7% on AIME 2026 and 90.2% on the German benchmark. It also scored 84.8% and 85.1% on GPQA Diamond and LiveCodeBench v6, respectively, competing against models like Qwen3.6-35B-A3B and Mistral Small 4 119B-A6B.
For specific tasks involving mathematics, programming, and long-context handling, Kolibri's performance is comparable to models with up to four times as many active parameters. The FP8 version requires approximately 78 GB of memory, while the BF16 version needs about 156 GB. Aleph Alpha recommends a minimum setup of one Nvidia H200, B200, or B300, or two A100 80 GB/H100 SXM5 units.
Additionally, since August 2, the European Commission has been authorized to impose fines on general-purpose AI model providers for non-compliance with the AI Act, with penalties reaching up to €15 million or 3% of global annual revenue.