Aleph Alpha released Kolibri on 3 October, putting a German-English model with 78.1 billion total parameters into public hands under the permissive Apache 2.0 licence. The move gives Europe something it has often demanded but rarely possessed: a substantial, locally trained model whose weights can be examined, downloaded and run outside an American or Chinese cloud. Kolibri is Aleph Alpha’s bid to turn “sovereign AI” from a policy slogan into an operational choice for ministries, manufacturers and aerospace groups that cannot casually export sensitive documents for inference.

Its test will be harder than its launch: whether European buyers find its performance, hardware demands and documentation sufficient to justify choosing it over better-known foreign rivals. The Heidelberg company describes Kolibri as a mixture-of-experts transformer with 78.1 billion parameters, though only 3.46 billion are activated for each token. That design aims to offer the breadth of a much larger model without paying its full inference cost on every request.

It supports German and English, tool calling and adjustable reasoning effort, functions aimed at document-heavy assistants and multi-step automated workflows.

A European training run, not merely a European wrapper Aleph Alpha says its teams built the system in Germany and trained it on infrastructure in Germany and Finland, under German and European law. The company used 768 Nvidia B200 graphics processors for three training stages: 20 trillion tokens of initial training, followed by 3.44 trillion tokens of mid-training and 200 billion tokens of long-context adaptation. That provenance matters because European public bodies and regulated companies have grown wary of systems whose data, model development and hosting arrangements cross several legal jurisdictions. Kolibri’s value proposition is not simply that it speaks German. Aleph Alpha is offering a chain of control over data curation, training and deployment, including the option to operate the model on infrastructure chosen by the customer.

“Compliance comes as an inherited property of the model,” Aleph Alpha says. That is a powerful commercial claim, but not a substitute for a buyer’s legal assessment. The EU’s rules for general-purpose AI models have applied since 2 August 2025, while the Commission’s voluntary General-Purpose AI Code of Practice is intended to help providers demonstrate compliance with obligations on transparency, copyright and safety.

Aleph Alpha says it built Kolibri with the AI Act, the code and the GDPR in mind, and its model card says the company is a code signatory. Deployment, personal-data handling and the eventual application still determine whether a particular use complies with the law. Kolibri’s documentation does offer elements procurement teams will want: a contact for rightsholders, disclosure about the training pipeline and a stated knowledge cutoff of 18 June 2026.

Yet it does not make the entire project open source. Apache 2.0 covers the published weights and configuration files; Aleph Alpha retains rights over its code, architecture, training methods and other unpublished artefacts. Open weights, in other words, are a meaningful degree of operational freedom, not full reproducibility.

Long context comes with a practical limit

The company advertises a context window of 1,048,576 tokens — enough in theory to ingest a vast case file, a technical archive or a long chain of records. But the model card recommends keeping requests at or below 262,144 tokens for serving efficiency and complex tasks. Kolibri was trained to that lower maximum length, a distinction that will matter more to engineers than to marketing departments.

Aleph Alpha says the model’s hybrid attention design helps restrain long-context computation: 40 of its 50 layers look through a 512-token sliding window, while every fifth layer processes the full context. The company also says the model needs roughly 78GB for its FP8 weights, with a B200 or B300 listed among recommended single-accelerator options. Sovereignty still requires expensive hardware, capable operators and a credible hosting plan.

The model’s bilingual emphasis is unusually concrete. Aleph Alpha says 21.3% of pre-training tokens were German, or about 4.3 trillion tokens, compared with roughly 62% English and 14% code. It says only 6% of training material was translated, arguing that translation can carry the institutional and cultural assumptions of English-language source material into German output.

A benchmark claim awaiting outside scrutiny In its own evaluations,

Aleph Alpha reports scores of 96.9 on English AIME 2025, 87.5 on its German version, 84.3 on GPQA Diamond and 85.9 on LiveCodeBench v6. It argues that Kolibri sits on the quality-versus-serving-cost frontier among selected open-weight systems, particularly for German and enterprise-oriented tasks. Those figures establish a promising starting point, not an independent verdict.

Benchmark outcomes depend on prompting, reasoning settings, evaluation harnesses and the competing models selected.