Aleph Alpha launches Kolibri, a 78‑billion open‑weight English‑German MoE model

On October 3, 2026, Aleph Alpha announced that it had released the full set of weights for its newly developed Kolibri model under the permissive Apache 2.0 licence, a move that coincided with Germany’s national day of reunification and highlighted the firm’s commitment to European AI sovereignty.

Kolibri is a mixture‑of‑experts (MoE) transformer designed to operate natively in both English and German. Although the architecture contains a total of 78.1 billion parameters, only roughly 3.46 billion of those are activated for any given token, a strategy that cuts inference compute while preserving the model’s overall capacity.

Technical specifications

The model features a native context window of approximately 262 000 tokens, a length that dwarfs the context sizes of most contemporary large language models and enables the processing of very long documents in a single pass.

For especially demanding scenarios, Aleph Alpha also supports an extended maximum context of 1 048 576 tokens, although the company advises users to keep contexts shorter in routine deployments to avoid sudden latency spikes.

Even with MoE‑based efficiency, the complete weight set remains substantial. MarkTechPost estimates that, when quantised to FP8 precision, the model occupies around 78 GB of storage, a size that can fit on a single Nvidia H200‑class accelerator or be spread across several 80 GB devices depending on the chosen runtime configuration.

Training data and methodology

Aleph Alpha reports that Kolibri was trained on roughly 20 trillion curated tokens, drawn from a larger pool of more than 200 trillion raw tokens. Of the curated corpus, 21.3 % consisted of German text while only 6 % were translations, underscoring a strong emphasis on native‑language material.

The training pipeline was engineered for resilience: checkpoints were generated automatically, evaluations were run about every hour, and the system was capable of continuing training through hardware or network failures without any manual intervention, according to the company’s statements.

Performance claims and regulatory stance

Aleph Alpha cites strong outcomes on both proprietary and public benchmarks, including head‑to‑head comparisons with models that activate a larger number of parameters per token. The firm stresses, however, that these results are vendor‑reported and still need independent verification.

Kolibri was built and trained in Germany and Finland with the EU AI Act, the General‑Purpose AI Code of Practice, GDPR, and copyright considerations in mind. While the open licence removes many access barriers, Aleph Alpha warns that the licence alone does not guarantee legal compliance for every downstream use case.

• Public‑sector applications such as tax administration and citizen services
• Industrial automation and predictive maintenance
• Automotive design, simulation and compliance checking
• Semiconductor chip design verification
• Aerospace systems engineering and safety analysis

The company positions Kolibri as a sovereign solution for regulated and mission‑critical workloads. By offering on‑premise deployment, Aleph Alpha aims to give organisations full control over their AI stack, a claim that resonates strongly with European data‑sovereignty policies.

For English‑speaking organisations, Kolibri’s bilingual capability provides a direct bridge to German‑language data without the latency and cost associated with translation pipelines. The expansive native context window also enables the processing of extensive contracts, technical specifications, or codebases in a single pass, streamlining compliance reviews and technical audits.

Because the model’s weights are openly available, firms can fine‑tune or audit the architecture internally, aligning AI behaviour with internal governance frameworks while avoiding dependence on external cloud providers.

In practice, early adopters in the finance sector have reported that Kolibri’s ability to ingest multi‑hundred‑kilobyte policy documents reduces manual review time by up to 40 %, while the model’s German fluency eliminates the need for separate translation services.

The release of Kolibri on Germany’s reunification day underscores Aleph Alpha’s strategic intent to champion a European‑first AI ecosystem, offering a high‑performance, open‑weight alternative to the dominant US‑based offerings.