Aleph Alpha Kolibri: Germany's Sovereign Open-Weight Model Explained

The Heidelberg company Aleph Alpha has released Kolibri, an English and German model with 78.1 billion total parameters, 3.46 billion active per token, and support for contexts up to 1,048,576 tokens. The weights are published under Apache 2.0, and it was released on 3 October 2026, the Day of German Unity.

This is a genuinely good day for European AI, and it is worth saying so plainly. A European company has published a capable model, with open weights, under a permissive licence, trained on European infrastructure, and designed from the start to run inside the customer's own estate. For organisations that have spent three years being told that serious AI means sending data to a handful of American providers, that is a meaningful change in the available options.

Kolibri is not available in the Owlpen platform, and we are not promising a date. More on that below. The rest of this article sets out what has actually been released, why the design choices matter, and what to weigh before treating it as a procurement answer.

What has been released

Kolibri is a mixture-of-experts transformer, an architecture that holds a large number of specialist sub-networks and routes each token to only a few of them. Aleph Alpha reports 384 experts across 50 layers, with six active per token plus one shared expert.

Size against running cost

The headline figures are 78.1 billion total parameters against 3.46 billion active, which is roughly 4.4% of the model doing the work on any given token. That is the central design decision. Capacity is stored in the full parameter count, while the cost of inference tracks the much smaller active count. Aleph Alpha reports serving 18 concurrent 256,000-token requests on two H100 GPUs, and decoding around 28% faster than a 123B baseline.

Context and formats

The context window reaches 1,048,576 tokens, built up in stages: pre-training at 16,384 tokens, mid-training at 64,000, and long-context adaptation to 262,144. Weights are published in FP8, a form of quantisation that puts the download at roughly 78 GB.

Reasoning effort and tools

The model exposes four reasoning effort levels (none, low, medium, and high) and supports tool calling, which is the practical requirement for agentic workflows rather than single-turn question answering.

Built for German, not translated through it

The most interesting engineering choice is linguistic rather than architectural. German accounts for 21.3% of the pre-training mix, around 4.3 trillion tokens, against roughly 62% English and 14% code. The tokeniser uses a 128,000-entry bilingual vocabulary designed to keep German compound words intact, and Aleph Alpha reports German compression of 4.90 bytes per token, the strongest in its stated comparison group.

That matters more than it sounds. Better compression on a language means fewer tokens for the same text, which lowers cost and leaves more of the context budget for actual content. It also means the model reasons in German rather than routing through English internally, which tends to show up in the quality of domain language: administrative, legal, and engineering German are not well served by a model that treats the language as a translation layer. Anyone who has watched a general-purpose model mangle German compound terminology in a regulatory document will recognise the problem being solved here.

The reported results support the ambition. Aleph Alpha publishes 96.9% on AIME 2025 in English and 87.5% in German, 84.3% and 81.3% on GPQA Diamond in English and German, 92.7% on HumanEval+, and 85.9% on LiveCodeBench v6. These are the vendor's own benchmark figures, and the company claims the model matches systems with up to four times as many active parameters on specialised tasks.

There is also serious work on knowing when to stay quiet. Aleph Alpha reports a non-hallucination rate of 44.0% on its AA-Omniscience evaluation, against 14.8% for the predecessor Kolibri Origin released in June, achieved through abstention training and an in-house method it calls the Merlin-Arthur protocol. Reducing hallucination by teaching a model to answer "I do not know" is unglamorous and exactly what regulated buyers need.

Why this is a win for Europe

The sovereignty argument is often made loosely. Here it rests on specifics. The model was built by teams in Germany and trained on infrastructure in Germany and Finland. The weights are published openly under Apache 2.0, which means an organisation can download them, inspect them, adapt them, and run them without asking permission or depending on a vendor's continued goodwill. Aleph Alpha states that the EU AI Act, the General-Purpose AI Code of Practice, and GDPR were considered from the start rather than retrofitted.

Combine that with on-premises deployment on two H100s and the practical position changes for a category of buyer that previously had no good answer: public administration, defence, aerospace, semiconductors, automotive, and industrial manufacturing, all of which Aleph Alpha has built sector-specific evaluation suites for. For those organisations the binding question has never been which model tops a leaderboard. It is who can see the data, which jurisdiction the processing sits in, and what happens if commercial terms change. An open-weight model you host yourself answers all three.

Releasing it on the Day of German Unity was a deliberate piece of symbolism, and the symbolism is earned. European AI has spent a long time being described as a regulatory bloc rather than a producer. A competitive open-weight model, published openly by a European company on European compute, is a straightforward counter-example, and the open licence means the benefit is not confined to Germany. Any European organisation, including UK ones, can use it on the same terms.

What to weigh realistically

Self-hosting is a capability, not a download

Open weights move the work rather than removing it. Running a 78 GB model in production means GPU capacity, serving infrastructure, monitoring, patching, evaluation, and the staff to operate all of it. For many organisations a hosted service remains cheaper and safer, and the honest comparison is total cost of ownership rather than licence cost.

The language scope is deliberate

Kolibri is specialised for English and German. That focus is the point, and it is also a limit. Organisations needing strong performance across French, Spanish, Italian, or Polish should test rather than assume, and the published figures do not speak to those languages.

Vendor figures need your own evidence

All the benchmark numbers above are Aleph Alpha's own, and the sector scores are drawn from its contextualised evaluations rather than independent testing. Published benchmarks have become a weaker guide to real-world behaviour across the industry generally. Any adoption decision should rest on measured accuracy on your own documents.

The corporate picture is moving

Aleph Alpha agreed a merger with the Canadian company Cohere, announced in April 2026 and signed as a definitive agreement in September, valuing the combined business at around $20 billion with backing reported at $600 million from the Schwarz Group. The combined company is reported as operating under the Cohere name with headquarters in Berlin and Toronto, subject to regulatory approval. Both companies present the combination as a sovereign alternative to the American incumbents. It is worth noting all the same that an organisation adopting Kolibri on sovereignty grounds should understand the corporate structure behind it. The Apache 2.0 licence is the real protection here, because published weights cannot be withdrawn.

Owlpen and Kolibri

Kolibri is not available in the Owlpen platform. Owlpen routes configured client workflows to a reviewed set of commercial models through vendor API access, with cost tracking, audit logs, and escalation paths attached. A self-hosted open-weight model is a different delivery shape: it needs hosting, serving, and operational ownership that sit outside the current platform scope.

That may change. Open-weight European models are a reasonable fit for clients with data residency constraints, and we expect demand for exactly that. If we add Kolibri, it will be after we have tested it against client material, established where the model is dependable, and confirmed how hosting and AI governance obligations are met. We are not setting a date.

In the meantime, organisations wanting to evaluate it can do so directly. The weights are on Hugging Face under Apache 2.0, and nothing about the licence requires a commercial relationship with anyone. Where a client wants help assessing whether a self-hosted model suits their circumstances, that is a conversation we are happy to have on its merits, including where the answer is that it does not.

Sources used

This article is based on Aleph Alpha's Kolibri launch post for architecture, training, benchmark, and deployment detail, the Kolibri-1 model listing for licence and distribution, and Sifted's reporting on the Cohere merger for the corporate position. All performance figures are those published by Aleph Alpha and have not been independently tested by Coaley Peak.

If you would like to discuss model selection, data residency, or the Owlpen platform in the context of your own workloads, contact us at enquiries@coaleypeak.co.uk or read more about the Owlpen platform.