Aleph Alpha releases Kolibri-1, a 78B German-English model with Apache 2.0 weights

What Aleph Alpha released

Aleph Alpha released Kolibri-1 on 3 October 2026, the Day of German Unity. In its announcement, the company describes Kolibri as an English-German mixture-of-experts Transformer that was built in Germany and trained on infrastructure in Germany and Finland. The model card on Hugging Face lists 78.1B total parameters and 3.46B active per token. Each of the 50 layers has 384 routed experts and one shared expert, and six routed experts are selected per token.

The model has an explicit reasoning mode with effort settings of low, medium, high or none, and it supports tool calling. The context reaches 1,048,576 tokens, but the card recommends 262,144 tokens or less for serving efficiency. The knowledge cutoff is 18 June 2026. Weights ship in FP8 with a footprint of about 78 GB. The listed minimum hardware is two A100 80 GB GPUs, two H100 SXM5 GPUs, or a single H200, B200 or B300. Serving requires the aleph-alpha-inference package, which provides a vLLM plugin and a container image.

What the Apache 2.0 license covers

The model card states that the Apache 2.0 license covers only the weights and configuration files. Code, architecture, parameter settings and training methods are excluded. In practice, companies can run, fine-tune and redistribute the weights commercially under Apache terms, but this is an open-weight release rather than a fully open-source one with training code. Aleph Alpha says it is a signatory of the EU General-Purpose AI Code of Practice and that Kolibri was designed with the EU AI Act and GDPR in mind.

The card describes a bilingual pretraining corpus of 20T tokens, followed by 3.44T tokens of mid-training and 201B tokens of long-context extension. Pretraining ran on 768 NVIDIA B200 GPUs for about 21 days, or 392,000 GPU hours, with total compute of 6.4e23 FLOPs and an estimated 950 MWh of energy, excluding fine-tuning and reinforcement learning.

How Aleph Alpha says Kolibri performs

In Aleph Alpha's own tables, Kolibri scores 75.5 overall in English, against 74.7 for Qwen3.5 35B-A3B, 73.0 for Nemotron 3 Super and 72.3 for GPT-OSS 120B. The dense Qwen3.8 27B leads at 80.2. In German, Kolibri scores 70.8 against 79.9 for Qwen3.8 27B. The card also lists 84.3 on GPQA Diamond, 96.9 on AIME 2025 and 66.4 on SWE-Bench Verified, and notes that dense models activate several times more parameters per token. The blog adds a training method called Merlin-Arthur, in which the model learns to abstain when the supplied documents do not support an answer.

What to watch next

Aleph Alpha says enterprise deployment and specialization go through its sales team, and pricing for that is not stated in the announcement. The benchmark figures are vendor results, so independent runs on German public-sector and industrial tasks will be the next check. Plans for other Kolibri sizes are not stated. The release is listed on the AI model release tracker, and more open-weight releases are collected under Open Source AI.

Questions people ask

Is Kolibri-1 open source?

Its weights and configuration files are released under Apache 2.0. The model card says code, architecture, parameter settings and training methods are not covered, so it is an open-weight model.

What hardware does Kolibri-1 need?

The FP8 weights take about 78 GB. Aleph Alpha lists a minimum of two A100 80 GB GPUs, two H100 SXM5 GPUs, or one H200, B200 or B300.

Which languages does Kolibri-1 support?

German and English. Aleph Alpha calls the two-language focus a deliberate choice of depth over breadth.

Sources

1. Kolibri Has Landed: A Sovereign Open-Weight Model (Aleph Alpha, 3 Oct 2026)
2. Kolibri-1 model card (Aleph Alpha Research on Hugging Face, 3 Oct 2026)